The full book

In Bandwidth, I called myself the Forrest Gump of fiber. In this book I'm more like the guy in the seat next to Forrest. I was there, but the bench wasn't mine.

Data centers were never my main business. Fiber was. But the two grew up together, and I spent thirty years at the place where they touched. At MFS, I worked for the company that built MAE-East, one of the first rooms where the Internet's networks met. At Level 3, I ran the network that included the data-center Gateways we built in most of the big cities; by March 2000 we had about 4 million square feet across thirty facilities. At Zayo, we connected our fiber to every carrier hotel and data center we could reach, and for a few years we owned a colocation business of our own, zColo. After Zayo was taken private in 2020, its new owners sold zColo, and $600 million of the proceeds went to investors. One of my friends, Steve Smith, ran Equinix for about a decade and then took my old chair at Zayo.

So this is less of a memoir than Bandwidth was. Most of the people in this story never worked for me, and I never worked for them. I never met Ellen Hancock or K.B. Chandrasekhar, whose Exodus Communications was the hosting company of the dot-com boom. Where I was in the room, I'll say so. Where I wasn't, I'll lean on the reporters who were. The Wall Street Journal, the New York Times, the Washington Post, BusinessWeek and Fortune wrote the first draft of this history, and I quote them heavily. When their accounts disagree with each other, or with what I wrote in Bandwidth, the footnotes say so.

What I can offer is the pattern. I've lived through one Boom, Bust and Resurgence cycle, and I wrote a book about it. The data-center industry has already been through one of its own. It is in the middle of a second, bigger one right now, and this time the stakes are measured in gigawatts.

A Disclosure, and an Experiment

One more thing belongs up front, in plain sight.

This book is an experiment with Claude, the artificial-intelligence model made by Anthropic. Claude did almost all of the research and almost all of the drafting. It read the period newspaper archives, the SEC filings and the trade press, pulled the quotes, did the arithmetic and wrote the first version of nearly every page you're about to read, under my direction. I chose the story, set the rules and argued with the drafts. Claude did the digging, and it did a lot of it. If Bandwidth was a book I wrote with a lot of help, this one is a book I directed with a great deal of help.

I set the rules the way I'd set them for any researcher. Every fact, number and quote has to trace to a source listed at the end of its chapter. Nothing goes in my mouth that I didn't live. Where a sentence needed my memory and the record didn't have it, the draft leaves a bracketed note for me to fill in rather than making something up. You'll see some of those notes in this edition. I left them in on purpose.

The experiment has a wrinkle that I'd be a poor student of conflicts of interest to ignore. Anthropic is also a subject of this book. Chapter VIII covers its rise and its initial public offering, which was still pending as this draft was finished. So a chapter about Anthropic was drafted, in part, by Anthropic's product. In Bandwidth, I quoted Jack Grubman's line that "what used to be a conflict is now a synergy." I'd like to hold this book to a stricter standard than Grubman held himself. In the Anthropic section, every number comes from the company's SEC filing or is labeled as press reporting, and I've tried to give the skeptics the same room as the bulls. You'll be the judge of whether that worked.

Finally, I'm publishing this as a first draft, in public, and I will revise it. Machines and authors both make mistakes, and the people who were actually in these rooms will know things the archives don't. If you were there and the record here is wrong, tell me. The second edition will be better for it, and the footnotes will say who set us straight.

In Piketon, Ohio, stands the former Portsmouth Gaseous Diffusion Plant, a uranium site built for the Cold War.

In March 2026, the Department of Energy announced a new use for it. It would host a data center of 10 gigawatts, along with up to 10 gigawatts of new power plants to feed it, 9.2 gigawatts of them burning natural gas. SoftBank's energy arm would partner with the local utility, AEP Ohio, and the government cited $33.3 billion of Japanese money for the gas plants under a trade agreement. Energy Secretary Chris Wright said the project would "ensure the United States wins the AI race." Five months later, OpenAI announced that it had secured 8 gigawatts at the Ohio campus. CNBC called it the world's largest planned AI campus and put the possible cost above $500 billion.

Half a trillion dollars, at one site. A year and a half earlier, that had been the price tag on the whole of Stargate, announced from a podium at the White House.

Then Ohio voted. On November 3, 2026, voters across ten states faced forty-one local ballot measures on data centers, and twenty-one of them were in Ohio. They went against the data centers in every one but a single town in North Dakota. Ohio elected a governor, Amy Acton, who had backed a "conditional moratorium" on new data centers. Two days later the Nasdaq had its worst day in nineteen months, and the index of chip stocks fell more than 12 percent.

A Cold War uranium plant turned into an AI factory. A trade agreement paying for gas turbines. A price tag of half a trillion dollars. And a vote of no from the neighbors. In 1999, the buildings at the ends of Level 3's fiber were real estate, and dull real estate at that. Nobody campaigned on them. Most people never noticed them at all, and most still don't. They are windowless, gray and set back from the road. They hum. They sit in office parks in Ashburn, in old freight terminals in Manhattan, in a former Electrolux factory in Memphis, in the farm fields of Loudoun County and in a county seat in Oregon. And every text message, video stream, bank transfer and chatbot answer you've ever sent or received passed through one of them.

This book is about those buildings, the people who built them, and the money that came and went with them.

Everything but the Computer

The building is older than the Internet.

In April 1964, IBM chartered a special train to carry several hundred writers up the Hudson to watch it announce the System/360. The machines rented for $2,700 to $115,000 a month, and every company that rented one had to build a room for it: raised floor, air-conditioning, power, guards. The industry came to call it the glass house. Within a year, the New York Times was writing about "the computer center as a public utility" and quoting an executive's prediction for the "data-center industry."

In August 1983, the Times took its readers into a cave in the Catskills, behind a seven-ton steel door, where a disaster-recovery company kept a computer room waiting, lit and empty, for a customer whose own glass house might burn down. It had a "raised floor, air-conditioning, telecommunications hookups - in short, everything but the computer."

Everything but the computer. That phrase describes the product the entire data-center industry sells to this day. Someone else builds the room. Someone else brings the power, the cooling and the connections to the outside world. You bring the machines, or you rent them by the hour. The industry has spent sixty years arguing about who should own that room, who should pay for it, and how big it should be. The answer has changed every decade or so, and every time it changed, fortunes were made and lost.A few terms recur throughout this book, and they are worth defining at the start, the way Bandwidth's first footnote defined the Internet Backbone. A data center is a building, or a part of one, designed to house computers: it supplies reliable power, usually with batteries and generators for backup, cooling to carry away the heat the machines produce, physical security and connections to outside networks. The "glass house" of Chapter I was a company's own data center. Colocation is renting space and power inside someone else's data center for your own equipment. In retail colocation, a customer rents a cage or a few racks, buys power by the circuit and pays to connect to other tenants; in wholesale colocation, a customer leases a large block of space with its own dedicated power, often a megawatt or more, for several years. A carrier hotel (or "telecom hotel") is a building, often an old telegraph, telephone or industrial building, where many network carriers rent space to house their equipment and connect to one another. Interconnection is the business of connecting those networks and customers to each other inside a building, either through a shared switch at an Internet exchange point or through a cross-connect, a physical cable between two tenants' equipment, which is the smallest product in the data-center business and one of the most profitable. Hyperscale describes the largest buildings and campuses, built for or by the handful of cloud and AI companies, and measured in tens or hundreds of megawatts and, in the AI era, in gigawatts. A megawatt is a million watts. A gigawatt is a thousand megawatts, a bit more than the output of the nuclear reactor at Three Mile Island that Chapter VIII brings back to life.

Two Cycles and a Prologue

In Bandwidth, I told the story of the fiber business through a single cycle: Boom, Bust, Phoenix and Reconsolidation. The Boom of the late 1990s built the networks. The Bust of the early 2000s wiped out over a trillion dollars of value. The Phoenix rose from the wreckage, as buyers with cash picked up hard assets for pennies on the dollar. And over the next two decades, the industry slowly reconsolidated into a handful of national platforms.

The data-center business ran the same cycle, on the same calendar, often with the same people. Then it started running it again.

So this book has three parts.

The first is a prologue cycle, from the 1960s to about 1990. Computing was centralized in the glass house, and entrepreneurs figured out that most companies didn't want to own one. Ross Perot ran other people's computers under contract. Time-sharing companies rented computer time by the hour, forty years before the cloud. Sam Wyly rode that idea to a Market Value larger than Western Union's and then rode it back down. It was the industry's first Boom and Bust, complete with forecasts that ran ahead of reality, acquisitions paid for in stock and a write-down of the computers themselves. Then the personal computer arrived and seemed to kill the glass house.

The second part is the first full cycle, from the mid-1990s to about 2022. The Internet pulled computing back into buildings. In the late 1990s, the hosting Gold Rush raised billions on forecasts. By 2001, the forecasts had collapsed and the floors were empty. Out of the wreckage rose Equinix and Digital Realty, which bought finished buildings cheap and filled them. Then the cloud arrived, and a handful of giant tenants came to control most of the demand. By the eve of ChatGPT, private capital had bought up much of what was left, and data centers had become a mature, consolidated asset class, priced like a bond.

The third part is the second cycle, which began on November 30, 2022, the day OpenAI released ChatGPT. It is still in its Boom phase as I write this. Nvidia became the most valuable company in the world, selling the chips. The cloud companies began spending more on buildings and equipment in a single year than the whole fiber industry spent in the three years of its Boom. Three Mile Island is being brought back to life to power one company's data centers. And the scarce thing this time is electricity.

Where Bandwidth ended with a question about the next cycle, this book ends inside it. The question I asked at the end of Bandwidth still stands: are we before midnight, or after?

The Survivors

The bigger the mess, the longer it takes to clean up. In Bandwidth, it took two decades for the fiber industry to reconsolidate into four national providers: AT&T, Verizon, Lumen and Zayo. The data-center business did its cleanup faster, and it ended with a different cast. By the time the AI Gold Rush began, four groups stood on top.

The first is Equinix. It was founded in 1998 by two veterans of a Palo Alto Internet exchange, Jay Adelson and Al Avery, on a bet that sounded too simple to work: don't run anyone's servers and don't sell anyone's bandwidth; own the room where the networks meet and charge them to connect to each other. Equinix went public in August 2000, five months after the Nasdaq peaked, and traded for 30 cents a share in 2002. A rescue from Singapore and a run of bargain buildings from the Bust brought it back. Under Steve Smith, a friend of mine who later took my old chair at Zayo, it grew into the meeting place of the Internet in most of the world's big cities, with 175 data centers in forty-four markets by 2017.

The second is Digital Realty. A real-estate fund called GI Partners bought twenty-five data centers when nobody wanted them, leased them to companies like IBM and Cisco, and in 2004 took most of them public as the first data-center real-estate investment trust. Digital Realty was the landlord that barely noticed the 2008 crisis. Between 2015 and 2020 it bought Telx, DuPont Fabros and Interxion, and went from an American landlord of big, powered buildings to a global company in both wholesale and retail.

The third is the cloud companies: Amazon, Microsoft, Alphabet and Meta. They began this story as tenants. When Amazon started renting computers by the hour in 2006, the data-center business gained the largest customers it had ever seen. Then those customers learned to build their own buildings, in river towns in Oregon and potato towns in Washington, and they shifted the balance of power from the people who own data centers to the people who fill them. In the AI era, they are the biggest builders in the history of the industry.

The fourth is private infrastructure capital: Blackstone, KKR and Global Infrastructure Partners, Marc Ganzi's DigitalBridge, Stonepeak, EQT, Brookfield, Macquarie and the rest. Some of these are the same pools of money that bought Zayo, and some of them bid against each other to do it. Between 2015 and 2022, they bought nearly everything Equinix and Digital Realty didn't, including four data-center companies worth about $46 billion in eleven months. To them, a data center leased to a cloud giant for fifteen years was a stream of cash from the best credit in the world.

Behind the four survivors comes a new crowd, the Gold Rush companies of the AI era. Nvidia sells the chips. Oracle, CoreWeave and a swarm of developers build and rent the capacity. OpenAI, Anthropic and Elon Musk's xAI are the customers whose commitments fill the order books. None of them has yet lived through a Bust. In Bandwidth, I wrote that the companies that failed were the ones that confused building with earning. Some of the companies in the last chapter of this book are building faster than anyone in history. We'll see which of them are earning.

What I Saw, and What I Didn't

I come to this story from the fiber side. I said so in the note at the front of this book, and I'll say it again here, because it shapes what follows. At MFS, my employer built MAE-East, one of the first places where the Internet's networks met. At Level 3, I ran the network when we built data centers in most of the big cities. At Zayo, we ran our fiber into every carrier hotel and data center we could reach, and for about eleven years we owned a colocation business, zColo, which our new owners sold in 2020.

That gave me a good view of the buildings from the outside, standing where the fiber came in the door. It didn't put me inside the boardrooms of Exodus or Equinix or Digital Realty. So for most of this story, I rely on the reporters who covered it, and on the companies' own filings. Where I was there, I'll tell you. Where I wasn't, I'll tell you that too.

The Chapters

Chapter I, "The Glass House," covers the 1960s through the 1980s. IBM rents its computers by the month and becomes the landlord of corporate computing. Ross Perot builds EDS by running other people's computers and sells it to General Motors for $2.55 billion. GE, Tymshare and a swarm of imitators rent computer time by the hour, and the time-sharing Boom ends with a beer mug crashing down on your fingers. Comdisco and SunGard sell companies a spare computer room in someone else's building. And in 1989, Kodak hands its data centers to IBM, asking only to "plug into a wall and have data come out."

Chapter II, "Where the Networks Meet," covers 1986 to 1997. The government steps back from running the Internet. MFS sets up MAE-East in a Virginia office building, and a Digital Equipment researcher puts an exchange in the basement of Palo Alto's old phone building. In New York and Los Angeles, a telegraph palace, a freight terminal and an office tower become carrier hotels. And UUNET and AOL plant the first seeds of Data Center Alley in the farm fields of Loudoun County. The building where networks meet turns out to be worth more than the networks.

Chapter III, "The Hosting Gold Rush," covers 1996 to 2000. Exodus grows from a Sunnyvale office with no ceiling to a Market Value of more than $20 billion. Equinix makes the opposite bet, on neutrality. Bernie Ebbers pays about $6 billion to win Digex, in what WorldCom's John Sidgmore later called "the stupidest deal of all time." The carriers pile in, including my own Level 3. And the first warning comes from the power company, when one hundred server farms ask PG&E for 1,200 megawatts.

Chapter IV, "Empty Floors," covers 2001 to 2003. Only 22 percent of the data-center space in North America is used or sold. Exodus files for bankruptcy, and its best thirty data centers sell for about 4 percent of its peak Market Value. Global Crossing, PSINet, Genuity and the rest follow. Danny Lewin of Akamai dies on American Airlines Flight 11. The customers stay in the buildings. Only the names on the deeds change.

Chapter V, "The Phoenix," covers 2002 to 2010. Equinix buys the guts of a finished San Jose data center for about $900,000. GI Partners turns twenty-five unwanted buildings into Digital Realty. Rob Roy turns a $930,000 Enron leftover into Switch. Savvis buys the remains of Exodus for $155 million. And the 2008 crisis proves that a building full of real tenants on real leases survives even when the banks don't.

Chapter VI, "The Cloud Eats the Building," covers 2006 to 2016. Amazon rents computers by the hour on a credit card, which is Tymshare all over again. Google, Microsoft and Facebook build their own data centers in places nobody in the hosting business had heard of. Loudoun County becomes Data Center Alley. A new breed of wholesale developer goes public. And a fiber company in Boulder, Colorado, decides it should own some buildings too. That company was mine, and that section is as close to a memoir as this book gets.

Chapter VII, "The Reconsolidation," covers 2015 to 2022. Digital Realty and Equinix buy their largest rivals on both sides of the Atlantic. The phone companies that bought data centers at the top of the last cycle sell them back out. And private capital buys nearly everything else, including Zayo and, later, zColo. On the eve of ChatGPT, a data center looks like a bond.

Chapter VIII, "The AI Gold Rush," covers 2022 to today. Nvidia's data-center revenue goes from about $15 billion to $193.7 billion in three years. Stargate promises half a trillion dollars from the White House. CoreWeave, a former crypto miner, becomes this cycle's Exodus, or maybe its Global Crossing. Elon Musk powers Colossus with gas turbines on trailers. The cloud companies outspend the fiber Boom. A shuttered nuclear reactor is brought back to serve one customer, the grid runs short of capacity, and voters push back. Anthropic files to go public. And the ghost of Andrew Odlyzko asks whether today's forecasts are any better than the ones that sank the last Boom.

Chapter IX, "Before Midnight," keeps score. It weighs the survivors and the newcomers against each other, lays out the ways this cycle might end, and offers the lessons of two cycles in the plainest terms I can find.

Your Story, Too

In Bandwidth, I wrote that the story was also yours, because Boom and Bust cycles repeat themselves. I wrote that the next one was already ignited, and that it entwined AI with other new technologies. I didn't know then how fast it would arrive, or that it would arrive in the buildings at the ends of the fiber.

The pattern in this book is older than the Internet. A real idea attracts a forecast. The forecast attracts capital. The capital builds ahead of the demand, on borrowed money, and the stock price runs ahead of the forecast. When the forecast comes down, the companies that borrowed against it go bankrupt. The buildings survive. Buyers with patience pick them up for a fraction of what they cost, fill them with tenants and make fortunes. Then the industry consolidates, the survivors get comfortable, and something new arrives to start the music again.

It happened to time-sharing in 1969. It happened to hosting in 2001. Whether it happens to AI is the question at the end of this book.

I can't tell you when the clock will strike midnight. Nobody can. Buffett's point was that the clocks at the ball have no hands. What I can do is show you the last two times the music played, who was dancing, who left early and who stayed too long. If you know the pattern, you'll recognize it when it comes around again.

It's coming around again right now.

On April 7, 1964, a special train pulled out of Grand Central Terminal carrying several hundred writers up the Hudson to Poughkeepsie. IBM had invited them to watch it announce a new family of computers. It did not trust a single room to hold the moment. That same day, more than 100,000 businessmen sat through simultaneous presentations in 165 cities, and the system was announced in more than eighty countries. Watson called it "the most important product announcement in the company's history."The New York Times account, which is unsigned, names him only as "Watson." IBM's chairman in 1964 was Thomas J. Watson Jr.; his father had died in 1956.

The product was the System/360. It came in six processor models, numbered 30 through 70. A customer could rent the smallest for $2,700 a month or the largest for $115,000 a month. IBM's Fred Brooks described bulk storage of 8 million characters. Deliveries were set for 1965 and 1966.

IBM's competitors read the next morning's papers and tried to look calm. RCA's E. S. McCollister told the New York Times something no executive should ever say about a rival's launch: "I was relieved in fact when I found out about it."

He shouldn't have been.

Those rental prices tell you what kind of business computing was in 1964. You didn't buy a computer. You rented one from IBM, by the month, and you built a room for it. The room had to be cooled and wired and guarded, because the machine inside it was the most expensive thing the company owned that didn't have a smokestack. The machines got bigger, the rooms got bigger, and over the next quarter century the corporate computer room became an institution, with its own staff and its own budget. The industry came to call it the glass house.I have not found a period newspaper article that uses the phrase "glass house" for the corporate computer room. The term is in common use in the industry's later histories, and it runs through this book as shorthand for the centralized, company-owned computer room.

This book is about the buildings that hold the world's computers. The story starts here, decades before anyone said "data center" with a straight face. Except they did. In 1965 the New York Times was already writing about the "data-center industry."

In this chapter we'll meet the people who figured out, in the 1960s, that most companies didn't want to own a glass house at all. Ross Perot built EDS by running other people's computers. Tymshare, General Electric and a swarm of imitators rented computer time by the hour, which is the cloud's business model forty years early. Sam Wyly rode that idea to a stock-market valuation larger than Western Union's, and then rode it back down. In the 1980s, Comdisco and SunGard sold companies a spare glass house in someone else's building, in case theirs burned down. Then the personal computer arrived and convinced everyone the glass house was finished.

It wasn't. It never is.

IBM—Thomas Watson Jr.#

For most of the glass-house era, the landlord was IBM.

The numbers are almost hard to believe. By 1968, IBM had built an estimated 70 percent of the computers in use in the United States. Computer makers had taken in more than $6 billion in 1967 from leasing and selling machines. And most of IBM's business came as rent. Customers signed up for a machine and paid for it every month, year after year, so each January the company started with most of the year already sold. Years later, the Wall Street Journal reported that under the old rental model IBM had "80% to 90% of the year's business locked up before the year even started." Tom Watson Jr. once made his annual sales quota on January 2.

That's the kind of business I spent my career chasing: recurring revenue. IBM had it in 1964, with a machine in a room.

The System/360 was designed to lock that model in for a generation. The six models shared one design, so a customer could start small and trade up without rewriting its software. Every trade up meant a bigger machine, a bigger monthly rent and, usually, a bigger room. The glass house grew with the computer.

IBM also learned early that renting someone time on a computer was a business in itself. Its Service Bureau Corporation ran about eighty data centers by 1965, processing other companies' work on IBM machines. By 1968, IBM's "Call 360" service linked customers in thirty-four cities to IBM computers in Los Angeles, San Francisco, Chicago, Cleveland, Philadelphia and New York. The minimum was $100 a month, not counting the terminal and the phone line.

Rivals hated it. IBM sold the computers, then rented time on the same computers in competition with its own customers. "Time-sharing gives IBM a foot in the customer's door," one of them told the Wall Street Journal in October 1968. That month, the Journal reported that the Justice Department was weighing a move to force IBM out of the time-sharing service business entirely.IBM's position in services was already constrained by its 1956 antitrust consent decree, which is why the Service Bureau Corporation operated as a separate unit. The Justice Department's 1968 deliberations came against the backdrop of the FCC's Computer Inquiry, opened in late 1966, which asked where computing ended and regulated communications began. Three weeks later, IBM moved its time-sharing business into a subsidiary to blunt the antitrust pressure.

Keep that tension in mind. The company that owns the machines also wants to rent them out, and its customers worry about competing with their landlord. That tension never goes away.

EDS—Ross Perot#

Ross Perot started Electronic Data Systems with $1,000. According to the New York Times, "He used the money to lease time on other companies' computers."The New York Times (June 29, 1984) is the source for both the $1,000 and the time-leasing origin. EDS was incorporated in 1962 (EDS Form 10-K, 2003). Perot had joined IBM in 1955 and quickly became a top salesman (Associated Press, July 9, 2019).

Read that again, because it is the whole business model of this book. Perot didn't own a computer. He rented time on someone else's machine and sold the results to customers who needed data processed and didn't want to build a glass house to do it. Over time EDS went further and ran whole data-processing operations for its customers under long contracts. Today we'd call it outsourcing. In the 1960s it was a Texan who understood that most companies wanted the answer, not the machine.

The answer paid well. EDS's earnings in 1968 were eighty-five times what they had been three years earlier.

The IPO

The Wall Street Journal summed up the offering in two sentences: "When he went public in September 1968, buyers snapped up 650,000 shares at $16.50 a share-or 118 times fiscal 1968 earnings. The stock soared to $40 by January." Perot kept 81 percent of the company.

One hundred eighteen times earnings. In 1968. For a company that ran other people's computers.

By March 1969 the stock traded at about 300 times earnings, and EDS had a Market Value "close to half a billion dollars." Its annual sales were $8 million.

The Cold Cheeseburger

On March 24, 1969, Perot was thirty-eight years old and on his way to Richard Nixon's first White House state dinner. Before he left, he announced the boldest move of his career. EDS would swap stock worth $65 a share for up to 82 percent of Collins Radio. Collins had $440 million in sales. EDS had $8 million.

This was the 1960s version of a move I watched many times during the Bandwidth Boom: use your richly valued stock as currency to buy a real company many times your size. Perot called it "the classic take-over-a real financial coup."

Arthur Collins did not see it that way. He refused to meet Perot, and his smoking rose to more than three packs a day. Perot was blunt about his target's founder. Over a cold cheeseburger, he told the Journal: "His job is to conduct the symphony, but all he does is play the drums, leaving no one in charge." The institutions that owned Collins stock, led by Chase Manhattan, sided with management. After five weeks, Perot withdrew. His explanation was pure arithmetic: "It just wasn't worth more than 1.5 for 1."

Sam Wyly, whom we'll meet shortly, had been approached to come in as a white knight. He "bowed out when his own company's stock fell." Hold that thought.

The $445 Million Day

On April 22, 1970, EDS stock fell 50 points, from 150 to 100, in a single day of over-the-counter trading. The New York Times headline the next morning was "Perot's Stock Falls $445-Million in Day." The arithmetic was simple. EDS had about 11 million shares outstanding, and Perot owned about 81 percent of them.

Let's do the rest of the math. At $150 a share, EDS had a Market Value of about $1.65 billion. By the close it was worth about $1.1 billion. EDS's secretary said he knew of no reason for the drop. Earnings were fine: nine-month earnings per share had risen to 42 cents from 17 cents. A broker offered the only explanation that ever matters in a falling market. Dealers held back because there "are so few buyers."These are prices on the over-the-counter market. EDS had about 11 million shares outstanding in 1970, against 650,000 sold in the 1968 offering. I have not compared the 1970 prices with the $16.50 offering price because my sources do not say whether the shares were split in between.

Perot was spending much of his time then on a private campaign for the return of American prisoners of war in Vietnam. Later that year he led a group that put $15 million into the Wall Street brokerage F.I. duPont, and a firm he controlled won the right to acquire duPont-Glore Forgan. In April 1971, the Journal reported another paper loss for Perot, this one $50 million. Wall Street was a costly detour. EDS itself kept hiring. In January 1970 the astronaut Frank Borman announced that he was leaving NASA to join a Perot firm.

General Motors

Fourteen years after the $445 million day, the top U.S. automaker walked in the door.

On June 29, 1984, the Wall Street Journal reported "GM to Acquire EDS in Transaction Valued at as Much as $2.55 Billion." EDS had about $800 million in revenue, roughly 1 percent of GM's. GM had more than $9 billion in cash. Perot and his family owned about 45 percent of EDS and would collect $900 million to $1.1 billion.GM offered EDS holders a choice: $44 in cash per share for about 58 million shares, or $35.20 in cash plus one-fifth of a share of a new GM Class E common stock, tied to EDS's performance. The all-cash value was $2.552 billion. GM guaranteed a minimum value for the Class E shares that rose to $125 at seven years. The Journal also reported that the deal would dilute GM's earnings by 99 cents a share for five years, and that the Class E structure might force the New York Stock Exchange to change its voting rules or delist GM. The Times described the price as $2.5 billion; I use the Journal's $2.55 billion, which matches the all-cash value and the October 1984 closing report. GM completed the deal that October.

The Journal soon reported that GM was considering moving 10,000 of its own employees onto the EDS payroll. That is outsourcing in its purest form. The top U.S. automaker decided that running its own computer rooms was not its business, so it bought the company that would run them and handed over the staff.

Perot had his own explanation of how he built the place. "Eagles don't flock," he said. "You have to find them one at a time."

Here's the peak-to-trough device run in reverse. In April 1970, EDS was worth about $1.65 billion at the open and $1.1 billion at the close. Fourteen years later, after enormous growth, GM paid about $2.55 billion. The stock market of 1970 had priced in a decade and a half of success before it happened. We'll see the same thing happen to every Gold Rush company in Chapter III.

Time-Sharing—Tymshare, General Electric and the Computer Utility#

On April 4, 1965, the New York Times opened a story on an executive in Garden City, Long Island, in an Oxford-gray suit. He sent his company's data to a computer forty miles away, at ITT's new Subscriber Data Processing Center in Paramus, New Jersey, and got his answers back. He didn't own the computer. He didn't need to see it.

The Times gave the idea a name: "the computer center as a public utility." Computer service centers had existed for about six years. The industry already had more than 600 firms. "They range from one-room loft shops to multimillion-dollar businesses," the paper reported. ITT's Robert A. Leonard predicted that time sharing and data communications would lift what he called the "data-center industry" from $300 million a year to more than $1 billion by 1970.

The technology behind that prediction was time-sharing. As I wrote in Bandwidth, John McCarthy of MIT envisioned time-sharing of computers, and in 1962 BBN demonstrated it with one operator in Washington, D.C., and two in Cambridge. Later that year BBN installed a time-sharing system at Massachusetts General Hospital, so that doctors and nurses could create and read patient records from terminals. For the first time, users could work with a computer in real time instead of submitting a stack of punched cards and waiting hours for a printout.

Time-sharing made the expensive machine in the glass house divisible. A big computer could serve dozens of users at once, each typing at a teletype terminal over a phone line, each believing the machine was theirs. Divide the machine, and you can rent out the pieces.

By the Hour

By 1968, customers paid "$10 to $20 an hour for computer time, depending on the service," according to Jerome B. Weiner of Mandate Systems, quoted in the Wall Street Journal. A typical customer had two or three teletype terminals, used them about two hours a day and spent about $600 a month per terminal, not counting the phone bill.

That's the line I want you to remember from this chapter. In 1968 you could rent a computer by the hour. You didn't buy it, you didn't house it, and you didn't cool it. You dialed in, you computed, and you paid for the time. When we get to Amazon Web Services in Chapter VI, you'll see the same idea come back with a credit card attached.

General Electric was the giant of the business. By late 1968, GE said it had more than 50,000 customers in seventy-five U.S. cities using terminals linked to thirty-one GE time-sharing computers. One GE 635 could serve up to about 800 customers, though no more than about 200 at once. A smaller machine gives a sense of the operator's economics. The GE 420, announced in June 1967, handled thirty users at once, leased for about $17,000 a month and sold for $760,000. GE predicted "that within five years, millions of persons will be using such systems."

The Frenzy

Wall Street fell hard for the computer utility. In 1968, Western Union and Computer Sciences planned to merge into one. Western Union's stock ran above 50 when the talks were announced. Computer Sciences peaked near 64 that June. The regulators were baffled. The FCC's Bernard Strassburg described the problem created by "this mating of computer and communications," and AT&T suspended a sale of its TWX service to Western Union over the deal. The Journal cited a forecast that 60 percent of the nation's computers would be linked to the communications system by 1970. By August 1968 the merger was off, and Western Union had fallen to about 35.

The forecasts got bigger every month. In October 1968, EDP Industry Report projected that time-sharing revenue would double 1967's $50 million in 1968 and reach $900 million by 1972. More than seventy companies offered the service. A trade group of eighteen to twenty members, the Time Sharing Venders Association, was forming. Tymshare's president, Tom O'Rourke, wanted IBM's Service Bureau divested: "We could compete on an equal footing."

And the futurists were out in force. Crystal-ball gazers, the Journal reported, predicted "housewives one day will send bank deposits and cash checks by punching data into home keyboard terminals." In February 1969, the New York Times headlined "Computer Time Sharing Grows Up." An Auerbach Corporation survey put the 1968 market at about $70 million, with forecasts of $140 million for 1969, $240 million for 1970 and $1 billion within five years. About thirty firms were in the field full-time in May 1968. By October there were more than fifty. GE's Information Service Department held about 40 percent of the market, IBM's Service Bureau about 19 percent, Call-A-Computer 7 percent, Com-Share 6 percent, Tymshare 5 percent and Allen-Babcock 3 percent.

The early believers were thrilled. "I told you so, I told you so," the Times had them saying. The goal, the paper reported, was "The computer in every kitchen, easily accessible and simple to use by every housewife and plumber."

They weren't wrong about the kitchen. They were thirty years early, and the computer ended up in the kitchen rather than at the other end of the phone line. Being early in a Boom costs the same as being wrong.

University Computing—Sam Wyly#

No company rode the computer-utility mania higher than University Computing Company of Dallas, and no one rode it harder than its chairman, Sam Wyly.

Metz called UCC "a leasing outfit in Dallas," and by 1968 it was thinking about networks as well as computers. The New York Times's Robert Metz opened his May 9, 1968, "Market Place" column with a line I wish I had written: "WHEN University Computing, a leasing outfit in Dallas thinks big, it thinks Texas big."

How big? UCC had $17 million of revenue in 1967. It tendered $44 a share for 750,000 shares of Western Union, 10 percent of the telegraph company and the legal limit. Western Union had $337 million of revenue. But UCC's stock sold at "well over 100 times 1967 earnings," and so the stock market valued the little Dallas leasing company above Western Union. A group of banks led by First National City lent UCC $44 million in Eurodollars, with $48 million more on standby, through a loophole in the Federal Reserve's Regulation U.Regulation U limited how much banks could lend against stock to finance purchases of stock. Borrowing in Eurodollars, outside the U.S. banking system, was the workaround (New York Times, May 9, 1968).

Think about that the way Metz's readers would have. A company with one-twentieth of Western Union's revenue was bidding for a piece of it, with borrowed money, because its stock was worth more. Business Week put its finger on the engine in May 1969: a "price/earnings ratio that bounced as high as 100 gave Wyly the leverage."

Wyly used it. In August 1968, UCC bid for Gulf Insurance. In January 1969, it closed a $240 million merger with the insurer. A computer-services company had bought an insurance company with paper. UCC also set up its own data-transmission business, and in November 1970 that unit took a complaint against AT&T to the FCC. Wyly was betting on a data network of his own, two decades before the commercial Internet. As we'll see in Chapter II, the idea was right. The timing and the balance sheet were not.

The Slip

The Boom ended the way they always do, slowly and then quickly.

On August 29, 1969, the day UCC was to list on the New York Stock Exchange, its president, Charles J. Wyly Jr., held to a forecast of about $15 million of 1969 net income, or $2.50 a share, on revenue of about $100 million.Charles J. Wyly Jr. was Sam Wyly's brother; the two started UCC together. UCC reported 1968 net income of $8.5 million, or $1.55 a share, on revenue of $57.1 million. A January 1969 Journal snippet refers to "Wyly, president of University Computing" without a first name. The numbers sounded great. But the forecast had already been cut twice that year, from $3.50 to $4 a share in February, then to $3 in June. Part of the blame went to a subsidiary, Computer Leasing Co. The rest went to "curtailed expansion of the company's computer network facilities, due to high interest rates."

Remember that phrase. A capital-hungry company building a network on borrowed money hits a rise in interest rates and has to stop building. That sentence will come back in this book more than once.

Then came the lawsuits. In February 1970, an $87 million suit was filed against LTV, LTV Aerospace and UCC over UCC's 1969 purchase of a company called Computer Technology from LTV. In May 1971, UCC reported first-quarter net income of $1.5 million and restated its profit for the year-earlier quarter.

In May 1972, Metz was back in the "Market Place" column with less flattering material. UCC had reported first-quarter net income of $11.8 million. But $10.6 million of it was capital gains realized at Gulf Insurance. The computer operations had an operating loss of $95,000, against operating income of $3.1 million a year earlier. Metz took UCC to task for comparing its results with "restated and knocked-down prior quarters." The stock closed at 24 3/8. The insurance company was carrying the computer company.

In August 1972, UCC agreed to sell a division to Harris-Intertype for $20 million and planned to take a $15 million capital loss. Then, in November, came the bottom. UCC lost $38.9 million in the third quarter of 1972 on revenue of $26.4 million. The charges included $21.2 million of losses on the sale of two subsidiaries, $7.2 million of other write-downs and a $9.7 million write-down of the company's own computer equipment.

That last item is the one I want you to notice. The computers had been the asset. Now they were a write-down. In Chapter IV we'll watch the same thing happen to raised floors and generators all over the country.

Here's the peak-to-trough. In 1968, the stock market valued UCC, with $17 million of revenue, above Western Union, at well over 100 times earnings. Four years later, UCC lost more in a single quarter than it took in.I have not found UCC's peak share price or a 1970 low. The stock closed at 24 3/8 on May 1, 1972.

UCC was the data-center industry's first mini Boom and Bust, and it followed the script Bandwidth readers will recognize: a real idea, a forecast that ran ahead of it, a stock price that ran ahead of the forecast, acquisitions paid for in paper, and then rising interest rates and write-downs.

The Shakeout—Keydata, GE and Tymshare#

UCC was not alone. The whole time-sharing industry went through the same cycle at the same time.

In August 1970, the New York Times ran a piece called "Keydata, Time-Sharing's Phoenix." Its writer, William D. Smith, captured the Bust better than anyone. Two years earlier, he wrote, firms had entered the business "at a rate of more than one a week" until about 200 were in the field. Three years earlier, if you mentioned time sharing in a Wall Street bar, bankers would offer to buy you a drink "and give him a million dollars." By 1970, "you are just as likely to have a beer mug come crashing down on the bar on top of your fingers."

The business had become "a revolving door of high promises and dismal performance."

The industry was then about $200 million a year, still forecast to reach $2 billion or more by 1975. A report from Quantum Science predicted that scientific time-sharing would stall at about $400 million a year after 1973, with thin margins. Compare that with the $900 million that EDP Industry Report had forecast for 1972, two years earlier. In the time-sharing Boom, as in every Boom since, the forecast was the asset.

Keydata itself was the survivor in the headline. Founded in Watertown, Massachusetts, by John T. Gilmore and Charles W. Adams, it launched in November 1965 as the first commercial time-sharing venture, was jammed with users by the summer of 1966 and had its first profitable month in May 1970. Four and a half years to a profitable month. That's what the survivors looked like.

The casualties came in clusters. Com-Share of Ann Arbor, with 6 percent of the market at the peak, was in financial distress by 1970 and needed new financing. Saul Steinberg's computer-leasing conglomerate Leasco reported a $30.8 million loss for its 1970 fiscal year. Computerworld said it plainly in June 1972: "For service companies 1971 was a 'shake-out' year."

General Electric

Even the market leader turned out to have been losing money all along. In August 1970, Honeywell's proxy for its purchase of GE's computer business disclosed GE's computer results for the first time. GE had lost $39.6 million in 1965 on revenue of $179.1 million, then $55.8 million in 1966, $42.5 million in 1967 and $11.4 million in 1968. It finally earned $4.7 million in 1969, on $411.6 million of revenue. Honeywell's chairman, James H. Binger, assured shareholders: "We are satisfied that the period of severe losses from these causes is over."

GE took about $237.5 million in Honeywell notes and stock and got out of building computers. But it kept something. GE held its U.S. and Canadian time-sharing services out of the deal.By 1985, GE's time-sharing business had become GE Information Services Co., known as Geisco, of Rockville, Maryland (Washington Post, October 21, 1985). GE had decided it didn't want to make the machines. It wanted to rent them.

Tymshare

Tymshare of Palo Alto came through the Bust cheaper, and alive. In September 1970, it sold 600,000 shares to the public at $6 each, raising $3.6 million. By late afternoon the shares were quoted at $6.25 bid and $6.75 asked. Burnham & Co. and Hambrecht & Quist were the underwriters, and the money went to repay debt and finance "increasing levels of receivables." Two years earlier, UCC had traded at more than 100 times earnings. Tymshare's IPO was a small, cheap offering in a slump.

The customers kept coming anyway. In November 1971, the New York Times described the controller of Inland Steel planning the company's five-year profit plan from his desk in Chicago. "A desk console in Mr. Borst's Chicago office is connected with a real-time computer at Tymshare in Palo Alto, Calif." The software came from On-Line Decisions, 30 percent owned by Inland Steel, which sold "what-if" simulations over time-sharing to Ralston Purina and Public Service Co. of New Mexico as well.

A steel company's finance chief in Chicago, running his numbers on a computer two thousand miles away that he would never see. That's the cloud, in 1971.

CompuServe—Jeff Wilkins#

The time-sharing company with the longest second act started as a way to use an insurance company's spare computer.

According to Elliot King's history of online journalism, Jeffrey Wilkins "founded Compu-Serv Networks, a computer time-sharing company, in 1969 primarily to make better use of the DEC PDP minicomputer used by his father-in-law's company, Golden United Life Insurance based in Columbus, Ohio."The CompuServe origin story comes from later books rather than period news coverage: Elliot King, Free for All (2010), p. 97; Sasser, Hart and Heskett's Harvard case collection (1991), p. 247; and Michael Banks, On the Way to the Web (2012), p. 15, which identifies Golden United Life as a subsidiary of Golden United Investments. H&R Block agreed in November 1979 to buy CompuServe for about $23 million in cash and stock, with completion set for early 1980 (Standard & Poor's, December 14, 1979). The insurer owned a computer it didn't use all the time. Wilkins sold the time it didn't use.

Hold on to that idea, because it reappears in this book under another name. A company builds computing capacity for its own peak, discovers that the machines are idle much of the day, and sells the leftover capacity to outsiders.

By 1985 the leftovers had become a consumer business. "H&R Block's Compuserve" competed with Reader's Digest's The Source, Dow Jones and Knight-Ridder's Viewtron, and in October 1985 GE joined them. GE Information Services, the descendant of GE's 1960s time-sharing business, launched a home service called GEnie at a flat $5 an hour during evenings, weekends and holidays. These services had more than half a million subscribers and about $70 million a year in revenue, though barely 100,000 regular users.

GEnie's manager, Bill Loudin, explained the economics to the Washington Post in two sentences. "We'll be using our excess capacity," he said. "Ninety percent of our computers just sit there at night and do nothing."

Twenty years before Amazon Web Services, GE was renting out its idle computers by the hour.

Comdisco—Ken Pontikes and SunGard#

By the early 1980s, the glass house had a new worry. What happens if it burns down?

The question created a new business, and two companies led it. One of them came out of the market for used IBM machines.

Comdisco

Kenneth N. Pontikes ran Comdisco, of Rosemont, Illinois, outside Chicago. In 1982 the New York Times called it "the largest independent leasing company dealing in I.B.M. computers." Comdisco's home market was used IBM equipment. IBM had long sold from a fixed price list that, one observer told the Times, "was like the Sears or Montgomery Ward catalogue."The speaker of the catalogue line and of "The steamroller is beginning to roll" is not identified in my notes on the article. Everyone knew IBM's price, so everyone knew how to undercut it. "They've basically been sitting ducks for everybody in the industry for years," Pontikes said.

That changed in 1982. In January, the government dropped its thirteen-year antitrust suit against IBM. IBM began to offer volume discounts and to negotiate "special bids." Analysts expected IBM's revenue to grow from more than $29 billion to about $50 billion in five years. "The steamroller is beginning to roll," the Times reported.

By then Comdisco had found a second business, built on the fear of disaster.

The Mountain

In August 1983, the New York Times took its readers inside a cave in the Catskills. The site had once been called Iron Mountain Atomic Storage. Behind a seven-ton steel door stood twelve buildings. One held a computer room that waited, lit and empty, with a "raised floor, air-conditioning, telecommunications hookups - in short, everything but the computer."

That room is the most important image in this chapter. Everything but the computer. A raised floor, cooling, power and telecom connections, ready for someone else's machines. Strip away the cave and the steel door, and you're looking at the product the entire data-center industry sells today.

The Times reported that the leaders in these "hot sites" were Comdisco Disaster Recovery Services, with four, and Sungard Services of Wayne, Pennsylvania, with three. Customers paid up to $7,500 a month for the right to move into a hot site if disaster struck. Empty "shells," with space but no equipment, cost $200 to $1,000 a month. Comdisco's president, Raymond R. Hipp, said the hot sites had been used for real twice. Customers drilled anyway, some flying as many as 150 people to Sungard's Philadelphia site to practice.

The sales pitch wrote itself. A University of Minnesota study estimated how long a company could survive without its computers, a figure it called the "maximum time to belly up." Banks could last two days. Insurers could last 5.6 days. Experts estimated that only 10 to 25 percent of heavily computer-dependent companies had adequate plans. "This is the place to come if you want to stay in business," the Times was told.

Here's why this matters to the rest of the book. For twenty years, the glass house had been something a company owned and kept on its own premises. Comdisco and SunGard sold companies a second glass house, in someone else's building, shared with other customers and paid for by the month. They sold insurance, but the product was real estate: a powered, cooled, connected room. The colocation industry of the 1990s grew from the same idea.

SunGard

SunGard Data Systems, the "Wayne, Pa., computer-service concern," went public on March 13, 1986. It sold 2.3 million shares at $11. The shares traded as high as $12.50 and closed at $12 over the counter.SunGard was created in 1983, when Sun Oil (Sun Co.) sold its disaster-recovery business and several investment-support businesses (SunGard investor presentation, 2004).

The Wall Street Journal mentioned it in a story about a different company. On the same day, a software company priced its shares at $21 and closed at $28. That was Microsoft.

I can't think of a better pair. The company that would put an operating system on every desk and the company that ran back-office computer rooms went public on the same day. We all know which one became bigger. But neither the desk nor the computer room was going away.

Comdisco, for its part, had cash to burn. In the October 1987 crash, its arbitrage unit, Comdisco Equities, lost $100 million, half its capital, before making back $20 million. Its head, Martin A. Weinstein, told the Times: "I was John Wayne and Rambo all at once." Making money from computer rooms and losing it in the stock market is a pattern we'll see again.

So is Comdisco. A dozen years later, when a startup called Equinix built its first Internet exchange in Ashburn, Virginia, Comdisco lent it part of the money for the build-out.The Equinix prospectus describes the Ashburn financing as Comdisco loans of $3 million "hard" and $4 million "soft," at 7.5 and 9 percent. Equinix accepted its first customers in Ashburn in July 1999. The company that rented out spare glass houses helped pay for one of the first neutral ones.

Dracula's Crypt

One more scene from the 1980s belongs here, because it points straight at Chapter II.

In January 1986, an entrepreneur named Paul Catalano opened DataPort, a three-acre vault 100 feet below the World Trade Center, in an old railroad terminus he leased from the Port Authority for forty years. He invested $3.5 million. Of the space, he said, "it smelled like Dracula's crypt. I fell for it right away." By August 1986, DataPort held 70,000 tapes for seventy-five clients. It had a hookup to Teleport's fiber network, and Catalano planned what he called "electronic vaulting," backing up data over the wire instead of by truck.

About 200 firms offered off-site storage then, with revenue of $100 million to $150 million that some thought could pass $1 billion in the 1990s. The competition was not impressed. Iron Mountain's Richard Reese called Catalano "a master of hype."

Maybe so. But look at what Catalano had: a secure vault, in lower Manhattan, wired to fiber. A decade later, buildings a few blocks away with the same three ingredients would become some of the most valuable real estate in the Internet business.

The Personal Computer—Katherine Hudson and the Kodak Deal#

While Comdisco and SunGard were selling backup glass houses, the main glass house was being emptied from the inside.

On November 23, 1987, the Wall Street Journal ran a front-page story headlined "Big Blues: Mainframe Slowdown And Stiff Competition Put Pressure on IBM." Mainframes supplied about 60 percent of IBM's profit. IBM held about 85 percent of the market for IBM-compatible "big iron." The paper's glossary defined a mainframe for its readers as "roughly the size of a refrigerator and typically costing more than $1 million."

The trend lines were brutal. The industry trade group CBEMA forecast that U.S. mainframe sales would grow only 3.3 percent a year from 1986 to 1996, against 8.8 percent for minicomputers and 14 percent for personal computers. It expected the PC market to pass the mainframe market in 1989 and to be more than twice its size by 1996. Users told the Journal that processing power could cost "as much as 150 times more on a mainframe than on a personal computer."

The customers were moving out. Hospital Corp. of America bought tens of millions of dollars of NCR minicomputers. The brokerage Edward D. Jones bought $30 million of DEC PCs and minicomputers for more than 1,300 branches. NBC's Olympics software cost about $30,000 to build on minicomputers, against more than $80,000 on a mainframe. "People can't even use the [mainframe] horsepower that's out there," said David Turner of National Advanced Systems. "I think you'll see more processing power deployed at the local level," said Thomas Cato of Hospital Corp.

I saw a small version of this from the bottom of the Bell System. As I wrote in Bandwidth, my first job out of college was at Illinois Bell, part of Ameritech. In Ameritech's planning groups I built Lotus 1-2-3 spreadsheets and Harvard Graphics presentations on a personal computer, while the thirty-year veterans I worked for stuck with their Texas Instruments calculators. They didn't know what to make of me. My dad spent his career as an engineer at AT&T. I spent my first job with a PC on my desk.

Cartons in the Hallway

Kodak's Katherine Hudson saw it more clearly than most, because she had to walk past it every day.

Hudson was Kodak's director of corporate information systems. For years she walked past a PC storeroom at Kodak's Rochester headquarters with cartons of equipment spilling out into the hallway. In 1989 she decided Kodak would stop running its own computers.

On July 26, 1989, the Associated Press reported that IBM "will run a large part of the data-processing operations of the Eastman Kodak company." IBM would build and operate a data center at Kodak's headquarters in Rochester, taking over work then spread across three Rochester centers and one in Windsor, Colorado. Terms were not disclosed. The Kodak employees who had run those centers would become IBM employees.The two Journal accounts disagree on the headcount. The July 26, 1989, item said 300 Kodak workers would become IBM employees. The August 15 page-one story said 275. UPI reported on July 25 that about 275 would be offered IBM jobs. I have used neither number in the text. Hudson hired Businessland to run Kodak's PCs. That fall, Kodak gave its voice and data network contract to Digital Equipment.The Wall Street Journal reported on October 26, 1989, that Digital Equipment had beaten IBM for Kodak's "voice and data communications needs," even though IBM held "an unusual contract to manage all of Eastman Kodak Co.'s data-processing needs." Kodak split its outsourcing: the data center to IBM, the network to DEC.

Hudson told the Journal she was looking at outsourcing software and telecommunications as well, and that more than 1,000 people could move off her payroll. She sent pink socks to the IBM team. Many of the Kodak employees moving to IBM put "Welcome to IBM Country" signs on their desks.

The Wall Street Journal put the story on page one in August 1989, under a headline that summed up the decade ahead: "Calculated Move: Computer Firms Find Service Is What Sells, Not Fancier Hardware." Hudson, the paper reported, "wants to be able to plug into a wall and have data come out," the way Kodak bought electricity from the power company and dial tone from the phone company.

Twenty-four years earlier, the Times had called it "the computer center as a public utility." Now a Kodak executive was saying the same thing about her own company's computers. The utility idea had come all the way around.

And Kodak was not alone. Hibernia Corp. signed a Kodak-like deal with IBM, and its spokesman called it a utility contract, "almost as if we were buying power or phone service." The federal Transportation Department already had Unisys running its Washington computer center. The research firm Input forecast that outsourcing of computer operations would grow from about $4 billion in 1988 to $11.9 billion in 1993, 20 percent a year.

The glass house wasn't dying. It was changing owners. The companies that had built their own computer rooms were handing them to someone else to run. That someone, in 1989, was still the company that built the computers. It wouldn't stay that way.

The Sword

The landlord itself was in trouble.

On December 6, 1989, the Wall Street Journal ran another "Big Blues" story on page one. IBM would take a $2.3 billion restructuring charge and cut 10,000 of its 216,000 U.S. jobs. Gartner estimated that rivals would take 17 percent of the $35 billion market for IBM-compatible mainframes in 1989 and 24 percent by 1991, where they had rarely held more than 10 percent. Nearly a third of the IBM customers surveyed had won mainframe discounts of 15 to 25 percent, and more than a third had won discounts above 25 percent.

The old rental model, with the year's business locked up by January, was gone. Customers had learned to negotiate. The Journal described how George DiNardo of Mellon Bank got a better price from IBM: by "taking a big Japanese sword (he actually keeps one in his office), waving it over his head and yelling the name of one of IBM's Japanese competitors."

Twenty-five years after the special train to Poughkeepsie, a customer was waving a sword at IBM.

The Dinosaur#

By 1990, the mainframe looked like a dinosaur. The money was moving to the desk. The PC market was passing the mainframe market. IBM was writing off billions and cutting jobs. The most forward-looking companies in America were handing their computer rooms to outsiders and asking to plug into a wall.

Step back, and the first thirty years of this story already contain the whole book in miniature.

Computing started in a room that one company owned. Entrepreneurs figured out that most companies didn't want the room, they wanted the answer. Perot sold the answer under contract. Tymshare and GE sold it by the hour. CompuServe and GEnie sold the leftover hours at night. Comdisco and SunGard sold a spare room in someone else's building, with "everything but the computer." Each idea came back, bigger, in the decades that followed.

The capital markets had already run their first experiment, too. Time-sharing had a Boom, with forecasts of $1 billion and 200 entrants and a stock in Dallas worth more than Western Union. It had a Bust, with a beer mug on your fingers, a $9.7 million write-down of the computers themselves and a long, quiet shakeout. The survivors, like Keydata and Tymshare, were still standing when the music stopped, and Tymshare went public cheap. I would see the same cycle in fiber. You'll see it twice more in this book, in data centers.

But something was missing from all of it. The time-sharing companies of the 1960s needed the phone network to reach their customers, and the phone lines were billed on top of the computer time. When Sam Wyly tried to build his own data network, he ended up complaining to the FCC. Was the computer utility of 1965 simply early? I think so. The network it needed did not yet exist.

By 1990, the PCs on millions of desks were starting to want to talk to each other, and to the computers in other companies' buildings. Nobody in the glass house was watching that. Nobody at Kodak or IBM, in the hot sites or in the cave in the Catskills, saw what the network was about to do to computing.

The network was about to arrive. And it would need a room.

At 8100 Boone Boulevard in Tysons Corner, Virginia, sits a small data center, across the street from AT&T Government Solutions. Nobody would slow down to look at it. According to the people who were there, this is where MFS set up MAE-East in 1992, with UUNET, PSI and Sprint-ICM as its founding networks.Three terms recur through the rest of this book. A Network Access Point (NAP) was the NSF's name for the four exchanges it designated in 1994 to carry traffic between commercial networks after NSFNET shut down. An Internet exchange point (IXP) is the general term for any place where networks swap traffic, usually across a shared switch; the NAPs, MAE-East, MAE-West and PAIX were all IXPs. A carrier hotel (or "telecom hotel") is a building, often an old telegraph, telephone or industrial building, where many carriers rent space to house their equipment and connect to each other. An exchange is the switch. A carrier hotel is the building it lives in. Over time the terms blurred, because the exchanges always ended up in the hotels. Local lore says the site was carrying around half of the Internet's traffic in the early 1990s. That claim comes hedged, with a "said to have" in front of it, and I'll treat it that way.The traffic-share claim appears in the Center for Land Use Interpretation's site entry for 8100 Boone Boulevard, which says the building is "said to have" carried around 50 percent of Internet traffic in the early 1990s (paraphrased from a fetch-tool summary). Loudoun County's economic-development timeline says half the world's traffic passed through MAE-East by 1997. That is promotional copy. I have not found a primary source for either claim.

Lore or not, here is what interests me about that building. In Chapter I, the personal computer emptied out the glass house. Computing left the raised-floor machine room and moved onto millions of desks. By 1990 the mainframe looked like a museum piece. But the desks wanted to talk to each other, and the networks that carried their conversations had to meet somewhere. Each meeting place was a room. Each room needed power, cooling, security and, above all, other networks. Without anyone planning it, the network was about to pull computing back into buildings.

This chapter is the story of those first rooms. The National Science Foundation stepped back from running the Internet and blessed four Network Access Points, one of them run by my employer. Scott Yeager, a data guy from MFS Houston, and Rick Adams of UUNET sparked MAE-East. A Digital Equipment researcher named Brian Reid, touring California's missions with his daughter, decided that bandwidth was the new railroad, and he put an exchange in the basement of Palo Alto's old phone building. Paul Vixie kept one of the Internet's root servers there. In New York and Los Angeles, landlords discovered that Western Union's palace, a Port Authority freight terminal and a downtown office tower had a second life. And in the farm fields of Loudoun County, UUNET and AOL planted the seeds of what Chapter VI will call Data Center Alley.

I had a front-row seat for part of this. From 1991 to 1996 I worked at MFS, which built MAE-East and much of the dial-up plumbing for UUNET and AOL. As I wrote in Bandwidth, I knew nothing about the Internet when I joined MFS. Neither did anyone else.

NSFNET—Stephen Wolff and the Four NAPs#

The Internet started as a government project, and for most of the 1980s the government ran its backbone. As I wrote in Bandwidth, the NSF launched NSFNET, which funded regional research networks that collectively became the first Internet backbone. Bill Schrader, who later founded PSINet, built the first NSFNET backbone linking the national supercomputer centers. In 1991 Al Gore championed a $1.75 billion spending package for his vision of an information superhighway. While others chased the money, the NSF allowed UUNET and a few rivals to interconnect. Commercialization was underway.

That created a problem the NSF had to solve. If commercial networks were going to carry the Internet, where would they hand traffic to one another? The government's answer was the Network Access Point, or NAP.Three terms recur through the rest of this book. A Network Access Point (NAP) was the NSF's name for the four exchanges it designated in 1994 to carry traffic between commercial networks after NSFNET shut down. An Internet exchange point (IXP) is the general term for any place where networks swap traffic, usually across a shared switch; the NAPs, MAE-East, MAE-West and PAIX were all IXPs. A carrier hotel (or "telecom hotel") is a building, often an old telegraph, telephone or industrial building, where many carriers rent space to house their equipment and connect to each other. An exchange is the switch. A carrier hotel is the building it lives in. Over time the terms blurred, because the exchanges always ended up in the hotels.

The NSF put the idea in writing on May 6, 1993, in Program Solicitation 93-52. A NAP, it said, was "a high speed network or switch to which a number of networks can be connected via routers for the purpose of traffic exchange." The solicitation named California, Chicago and New York City as priority locations. Washington, D.C., made only the second-tier list of "desirable" sites. Two lines in the document mattered more than the rest. Networks that attached would pay "both an initial and an annual fee," an early ancestor of the cross-connect charge. And the traffic would be open to all comers: it "will not be restricted to that which is in support of research and education." With that clause, commerce was in.The solicitation set expected funding for all its awards, including the routing arbiter and the very-high-speed backbone, at "approximately $18,000,000 per year." That is a different measure from the Times's "$12 million in annual subsidies" for the backbone being retired. Proposals were due August 17, 1993. Backbone providers got no subsidy at all: "costs of operation of the NSPs will be recovered from users."

On March 2, 1994, Stephen Wolff, the NSFNET program director, posted the winners to the com-priv mailing list. Under Program Solicitation 93-52, the NSF picked MFS Datanet to run a NAP in Washington, D.C. Sprint got New York, though its NAP sat across the river in Pennsauken, New Jersey. Bellcore would manage two more: one in California, operated by Pacific Bell, and one in Chicago, operated by Ameritech. The reaction was immediate. Karl Denninger, who ran a Chicago network, asked the list what this meant for ANS, the nonprofit backbone operator, whose salespeople had been confident of winning most of the NAPs. They won none.

The money involved was small. The NSF award record for the MFS Datanet NAP lists a budget period from July 1994 to June 1996 and $211,587 actually obligated, against an estimated total of $664,966. The principal investigator was Steve Feldman of MFS Datanet in San Jose. The whole federal backbone subsidy that was ending amounted to "an estimated $12 million in annual subsidies," according to the New York Times. Hold that number in your head. In a couple of years, the private companies sitting at these meeting points would trade for billions.

The handoff took longer than planned. In its October 24, 1994, article "U.S. Begins Privatizing Internet's Operations," the Times reported that the plan had been announced in 1992 and was running late. On October 31, the regional networks were to be cut off from the NSFNET backbone and connected to "four new commercial network hubs, in San Francisco, Chicago, Washington and Pennsauken, N.J." The NSF's Don Mitchell was candid: "Quite frankly, this has taken longer than expected."

The Times described three of the hubs as the "primary" NAPs: Pacific Bell, Ameritech and Sprint. The fourth, in Washington, was "operated by Metropolitan Fiber Systems Inc." The San Francisco and Chicago NAPs used ATM, a switching technology that had never been tried at Internet scale, and they were lagging in signing up customers. ANS's Jordan Becker warned that, if the cutover went badly, "you could have islands of connectivity."

The phone companies built the NAPs the way phone companies build things: on new, elaborate technology, at their own pace. The Washington NAP was different. It was already running.

MAE-East—Scott Yeager, Rick Adams and MFS Datanet#

The NSF did not invent the Washington exchange. It adopted one.

John Curran, a BBN technology chief in the 1990s, set the record straight years later on an Internet history mailing list. MAE-East, he wrote, had been set up at MFS in Tysons Corner in 1992 by UUNET, PSI and Sprint-ICM, before the NSF got involved. The NSF's solicitation then proposed a no-cost NAP model based on MAE-East, and MAE-East was named one of the four NAPs while it was already in operation.Insiders disagree about how long MAE-East predated the NAPs. Bill Woodcock says about a year; Scott Bradner, in the same 2021 mailing-list thread, says several years. Curran, Woodcock and the CLUI entry all put MAE-East's founding at 1992.

How did a metro fiber company end up running the Internet's most important switch? As I wrote in Bandwidth, the spark came from Houston. When MFS acquired Network Communications in 1989, its cofounder, Scott Yeager, became the city manager of MFS Houston. Yeager developed a new kind of data service using customer equipment that ran on Ethernet, the same protocol that connected computers inside an office. He shared the idea with Rick Adams of UUNET. That conversation sparked MAE-East. It also led MFS to create a new business unit, MFS Datanet, focused on data. Bill Woodcock of Packet Clearing House, another veteran, has named Steve Feldman as MAE's principal engineer.

The mechanics were simple. MFS already had fiber running into the buildings where the networks were. Put an Ethernet switch in one room. Let every network run a line to it. Each network plugs in a router, and traffic flows between them across the switch. What the networks owed each other was a separate question, and in 1992 the answer was nothing.

John Sidgmore, who ran UUNET and later became my boss, remembered the founding ethic. In 1997 he told the Washington Post how it worked four years earlier: "Instead of creating billing systems, we said, 'Let's just interconnect the network and not charge one another.'" A handshake among a handful of small networks became the Internet's first business model for interconnection. It was called peering.

A story repeated online says MAE-East's first switch sat in a parking garage. I could not find it in any primary source, so I'm not telling it here as fact.

What I do remember from those years is the culture. As I wrote in Bandwidth, Kevin O'Hara hired me in 1991 to lead MFS's metro network expansions, after his predecessor handed him my résumé with the advice that "The last thing MFS needs is a Bell head." MFS Development was raw, blue-collar and fearless. We went from fourteen networks at the end of 1993 to forty-two U.S. markets by 1994. MFS had three business units: MFS Telecom for fiber, MFS Intelenet for voice and MFS Datanet for data. I was building the roads. Datanet was building the intersections. I didn't fully appreciate at the time which of the two would turn out to be more valuable.

I do remember the day the Internet showed up in our executive suite. Jim Crowe, our chairman and CEO, gathered his senior team in MFS's twenty-story blue glass building in Oak Brook, outside Chicago. Instead of clicking through slides, he went online in front of us. "The Internet is going to be a powerful force in all of business," Crowe said. Then he asked who had heard of Yahoo. Eight hands went up. Mine was not one of them.

MAE-West and the Crunch

MFS ran a second exchange in San Jose, MAE-West. Feldman's NSF award was filed from MFS Datanet's San Jose office. By the second half of the decade the two MAEs anchored the coasts: MAE-East in Tysons Corner, MAE-West in San Jose.

MAE-West lived in a fifteen-story office and telecom tower at 55 South Market Street. Years later, the building's owner told its history in a prospectus: "a predecessor to Verizon Communications installed one of the nation's first internet network access points, or NAP, in the property known as the MAE West." An affiliate of the Carlyle Group bought the tower on February 2, 2000, five weeks before the Nasdaq peaked. It became the first property of a company called CoreSite.

By 1996 they were overwhelmed. In August of that year, in an article headlined "Web Crunch: Slow Crawl on the Internet," the Wall Street Journal described a major interconnect in San Jose peaking above 95 megabits per second against 100 megabits of capacity, until a new switch eased the jam. The Journal didn't name it, but I'd bet on MAE-West. At peak hours, according to University of Michigan researchers, backbones were losing more than 10 percent of their packets. MCI tripled its backbone to 155 megabits per second and planned 622 by year end.

The equipment couldn't keep up either. "You can't just buy the next bigger [router] box, because there is no next bigger box," said BBN's John Curran. A Cisco executive, Peter Alexander, went further: "the Internet is going to become the CB radio of the '90s."

By then the NSF had shut down its backbone, and the business had passed mainly to seven companies: ANS (now owned by AOL), AGIS, BBN, MCI, PSINet, Sprint and UUNET. UUNET, by then, was owned by MFS. The big providers would now exchange traffic only with peers that had a minimum number of high-speed lines.

The handshake didn't last. In May 1997, the Washington Post reported that UUNET, by then owned by WorldCom, would start charging some networks for traffic it once exchanged for free. UUNET had spent more than $100 million on its own fiber. Most of the roughly three thousand Internet service providers leased theirs. Sidgmore, the same man who had remembered the free handshake, now explained why it had to end: "when [ISPs] say they want to 'peer,' they're saying they want to use your network for free." ANS was screening its peers too. Vint Cerf, by then at MCI, called it "purely a business situation."

The big networks were learning that a meeting place was worth fighting over. MAE-East's run in Tysons had a few years left. According to Loudoun County, it moved to Loudoun in 1998.Loudoun County's timeline says MAE-East relocated to Loudoun in 1998. The Washington Post's July 1999 AboveNet article still places MAE-East in Tysons Corner, operated by MCI WorldCom, and says AboveNet had an operation "in the same Tysons building." The two may describe different parts of a distributed exchange. But the lesson was already clear. Whoever controlled the room where networks met controlled a choke point on the Internet.

PAIX—Brian Reid, Paul Vixie and Digital Equipment#

If MAE-East was the exchange a fiber company built, the Palo Alto Internet Exchange was the one a computer company built. That distinction would matter a great deal in the next chapter.

The story as I told it in Bandwidth begins in 1991, when PSINet and UUNET cofounded a trade association of Internet providers. The association set up an early exchange point, where providers swapped traffic through a router operated by PSINet. That router moved from its original home in Santa Clara to Palo Alto, and the site became known as the Palo Alto Internet eXchange, or PAIX. Paul Vixie, the Internet pioneer whose work underpins the Domain Name System, was among its cofounders.Bandwidth and the 1997 New York Times account emphasize different things. Bandwidth describes PAIX as the descendant of the PSINet-operated trade-association router, moved from Santa Clara to Palo Alto, and names Paul Vixie as a cofounder. The Times credits Digital's Brian Reid with the vision and describes the exchange as Digital's facility. The two accounts are compatible if the association's router moved into Digital's building, but I have not confirmed it. The Times says the building was full in July 1997, ten months after opening, which implies an opening around September 1996; Chapter III's sources date Digital's commercial opening to July 1996. A December 9, 1997, Times correction noted that the missions Reid toured date from the eighteenth century, not the sixteenth.

The New York Times told the building's side of the story. In its December 8, 1997, article "Old Man Bandwidth; Will Commerce Flourish Where Rivers of Wire Converge?" John Markoff explained that the NAPs, "established in the early 1990's," had developed "a growing traffic jam" by 1995. That jam gave rise to the idea of a private exchange that bypassed them.

The man with the idea was Brian Reid, a Digital Equipment researcher and Arpanet-era pioneer. In 1994 Reid was touring California's Spanish missions with his fourth-grade daughter. That trip, according to the Times, is where the idea came to him. "Bandwidth in the late 1990's is important for commerce in the same way that railroads were important in the 1890's," he told the Times.

I'll admit I'm partial to a man who sees the world in bandwidth.

Digital found its junction in an unglamorous spot. The city of Palo Alto finished a $2 million, fifteen-mile fiber ring that, in the Times's words, "culminates in Digital's new Internet Exchange." The exchange itself sat "in the basement of the building that was once the local telephone company's switching office, just half a block off University Avenue." Four phone companies brought in 14.9 gigabits per second of capacity. The building was 100 percent occupied ten months after it opened. Thirty-five Internet providers had moved in, including @Home, WebTV, Deutsche Telekom, Hongkong Telecom, IIJ and UUNET. WebTV's Steve Perlman put it in Reid's terms: "It's like being right on top of the nexus of five major railroads."

Digital pitched the exchange as "a neutral meeting ground." No carrier owned it. More than half the world's Internet traffic was said to pass through Northern California. To prove the exchange could take the load, Digital used the AltaVista search engine, then handling twenty-five million queries a day, as what the Times called an "attractive nuisance."

Vixie was the exchange's most famous resident. Since about 1988, he had run Root Server F, one of the thirteen computers at the top of the Internet's address book. In September 1997 the Wall Street Journal asked, in a headline, "Is the Internet Outgrowing Its Volunteer Traffic Cops?" Thirteen root servers were kept by eight volunteers. Vixie ran his through the nonprofit Internet Software Consortium in Woodside. The rumor in the industry said Root Server F lived in somebody's basement. Vixie set the record straight: "I built the hardware ... in my basement, but that particular basement had no Internet connection." The server lived "at the Digital Internet Exchange, a facility maintained by Digital Equipment Corp. that provides high-end switching services for Internet service providers."

"This is very much a public-trust kind of thing," Vixie said of the job.

Think about what that means. One of the master directories of the Internet sat in the basement of an old phone building in Palo Alto, kept there because that was where everyone else's networks were. The machine room had come back. It had just changed landlords.

As I wrote in Bandwidth, Digital didn't keep PAIX. AboveNet bought it, and when Metromedia Fiber Network acquired AboveNet in 1999, PAIX came along. Vixie served as CTO of the combined company from 1999 to 2001. The people who had run that basement for Digital will return in Chapter III, with a different idea about what a neutral meeting ground was worth.

The Carrier Hotels—60 Hudson, 32 Avenue of the Americas, 111 Eighth Avenue and One Wilshire#

The exchanges needed buildings, and the buildings that worked best were old ones. A thread on an Internet history mailing list once passed along an observation from Hunter Newby: everything is where it is because it was there before. The networks of the 1990s followed the cables of the 1920s, which followed the telegraph.

60 Hudson Street

Start with Western Union. As I wrote in Bandwidth, Western Union was an early backer of Teleport, MFS's archrival in New York, and then lost interest and sold its 40 percent stake to Merrill Lynch for $4.7 million. Teleport later sold to AT&T for $11.3 billion. Western Union's timing was poor in fiber. Its taste in buildings was better.

The Western Union Building at 60 Hudson Street is a twenty-five-story landmark, built in the 1920s in what was then a dairy-and-warehouse district and tied to the undersea cables. By the 1980s, as the New York Times put it, Western Union "is no more." The building was still there, and so were the cables.

By 1995 the leasing agents knew what they had. The Times described 60 Hudson that summer as "the telecommunications industry building at Thomas Street." Audrey Novoa of Williams Real Estate was courting small telecom tenants, and she wasn't picky about their balance sheets: "So if they have lousy credit we'll rent them four walls, with minimum buildout."

Four walls and a riser to the cables turned out to be a business. By 1999, dozens of phone and Internet companies had moved equipment into 60 Hudson. Its owner, Hudson Telegraph Associates, whose general partner was Ken Carmel, was planning a new electrical substation. The Times called the building "one of the telecommunications hubs of the East Coast, a major Internet node and a passageway for a great deal of trans-Atlantic phone traffic."

The neighbors had a different description. TriBeCa loft dwellers now lived beside a building full of generators and cooling equipment. "The noise is just unbelievable," said Bruce Ehrmann, who lived nearby.

32 Avenue of the Americas

A few blocks away stood AT&T's Long Distance Building at 32 Avenue of the Americas. In 1932, every overseas call to or from the continent ran through it. The building ran to 1.16 million square feet.

In December 1999, just past the period this chapter covers, the Rudin family took title and planned to spend about $250 million to buy and renovate it. Wells Fargo lent $168 million. The Rudins rebranded it the "New York Global Connectivity Center." The roof would get fourteen generators and ten cooling units. Down in the subbasement vault, John Gilbert III tried to count the four-inch copper cables. "We stopped counting at 600," he said.

The neighbors were already fighting 60 Hudson, which the Times called "another mesalike Art Deco landmark by Mr. Walker." The same architect had designed both. Councilwoman Kathryn Freed didn't want a second one making noise: "One is bad enough."

111 Eighth Avenue

Then there was the biggest of them all. 111 Eighth Avenue was the Port Authority's Inland Freight Terminal, built in the early 1930s. It covers the entire block from 15th to 16th Street, between Eighth and Ninth Avenues, with 2.3 million square feet of space.The Wall Street Journal (March 8, 2000) gives 111 Eighth Avenue as 2.6 million square feet. The New York Times (January 25, 1998) gives 2.3 million. I've used the Times figure, which is the earlier and more detailed account. Its floors are 175,000 square feet each, with fifteen-foot ceilings. Its truck elevators could lift a twenty-ton, thirty-foot truck.

The freight was gone, but the elevators found new cargo. In January 1998, the Times reported that telecom tenants were using them to hoist battery clusters "so heavy that the building floors must be reinforced with steel." Seven diesel generators sat on the roof with seven days of fuel. Rents ran in the high teens to low twenties per square foot, against about $40 in midtown.

Taconic Investment Partners, formed by Charles Bendit and Paul Pariser on September 1, 1997, teamed up with Blackacre and Goldman Sachs to buy four Sylvan Lawrence buildings, including 111 Eighth, for $387.5 million. Bendit called 111 Eighth "a sleek battleship." The Wall Street Journal later reported that when Taconic bought the building it "inherited three telecom tenants," and that five carriers already had fiber inside. So Taconic went looking for more telecom and Web-hosting tenants. A developer explained the logic to the Journal in a single line: "If no fiber cables go within five miles of a center, it doesn't do the tenants much good."

One Wilshire

On the West Coast, the equivalent was One Wilshire in downtown Los Angeles. By 1999, telecom companies had leased 1.1 million square feet of downtown Los Angeles office space, about 3 percent of the stock. They made up two-thirds of all downtown leasing that year. In August 2000, the Journal explained why: "Companies are clustering downtown to be near One Wilshire, a key switching center."CoreSite, the best-known colocation operator at One Wilshire, was a tenant there, not the owner. Its 2010 prospectus describes a lease that began in August 2007 and covered 172,970 square feet of the 664,108-square-foot, thirty-story tower, with more than three hundred customers and service providers interconnecting in the building. The carriers paid about $2 per square foot per month, often a third above market, for buildings others considered decrepit.

The city wasn't thrilled. It floated the idea of a "telecommunications switching district." Councilman Nick Pacheco's complaint summed up the problem: in these buildings, "only two or three people are working."

That was the whole point. A carrier hotel is a building for machines.Three terms recur through the rest of this book. A Network Access Point (NAP) was the NSF's name for the four exchanges it designated in 1994 to carry traffic between commercial networks after NSFNET shut down. An Internet exchange point (IXP) is the general term for any place where networks swap traffic, usually across a shared switch; the NAPs, MAE-East, MAE-West and PAIX were all IXPs. A carrier hotel (or "telecom hotel") is a building, often an old telegraph, telephone or industrial building, where many carriers rent space to house their equipment and connect to each other. An exchange is the switch. A carrier hotel is the building it lives in. Over time the terms blurred, because the exchanges always ended up in the hotels. The Journal would later describe the breed as "windowless buildings with reinforced floors, super electrical feeds and wiring to the max." The architecture hardly mattered. The other tenants did. Every network that moved in made the building more attractive to the next one. Telegraph palaces, freight terminals and phone switching offices had what no new building could offer: the cables were already there.

Northern Virginia—UUNET, AOL and the Internet Capital#

The meeting places of the East Coast settled in Northern Virginia for the same reason. That was where the networks already were.

Start with the companies. As I wrote in Bandwidth, Rick Adams joined the Center for Seismic Studies in Northern Virginia in 1982, where he worked on nuclear test-ban verification and became an ARPANET user. In 1987 he convinced a nonprofit of Unix programmers to put up $250,000 for Cisco hardware, and UUNET was born. The Washington Post says UUNET started "in a borrowed office in Arlington County in 1987." In 1989 Adams moved the hub into his home. UUNET had four employees on its first day as a business. "If people were begging you to take their money for this, that's the business I want to be in," Adams said.

America Online grew up a few miles away. In August 1994, the Post called it "the Vienna-based on-line computer service." AOL expected to reach a million households by the end of that summer, after tripling its base in a year. Subscribers dialed in and received information "sent over telephone lines from central AOL computers." The whole industry was adding about twenty-two thousand new customers a week.

Central computers. In the age of the personal computer, AOL's customers were connecting their PCs to someone else's machine room. The glass house was coming back through the phone line.

AOL kept growing, and it moved. By 1996 it had relocated its headquarters from Vienna to Loudoun County, with eleven hundred people at Dulles. By December 1997, according to the Wall Street Journal, it had more than nine million members, and its North American support calls were routed through "a routing system in Virginia."

Then UUNET followed. On Christmas Eve 1997, the Washington Post ran the news on its front page: UUNET planned a Loudoun headquarters with more than three thousand workers by 2000, on a roughly 400-acre site just north of Dulles. The land was the "Glenwood" farm near the Dulles Greenway, where Jack Kent Cooke had once hoped to build a stadium for the Redskins. Bernie Ebbers had walked the candidate sites himself. UUNET's work force had grown from 450 to nearly one thousand in a single year, spread over four buildings in Fairfax and Tysons. Virginia offered $1 million to $10 million in grants. The Post said the campus would "become the control center for much of the world's online traffic."

Loudoun's leaders could hardly believe their luck. "It's exciting to think we could become the Internet capital of the world," said Supervisor David McWatters.

He undersold it.

Others saw the same pull. In July 1999, the Post reported that AboveNet, the San Jose exchange operator, had agreed to buy a nine-acre tract in Reston for a 250,000-square-foot "Internet Service Exchange," where the servers of Internet providers and content companies "can sit next to each other." AboveNet already had an operation in the Tysons building that housed MAE-East. Its executive vice president, Stephen Belomy, explained the choice of Northern Virginia: "topologically, it's central to the Internet." The area had "a lot of networks that have some kind of presence in the area." The new building would house fewer than one hundred people, where an office of the same size would hold a thousand or more. It would cost "hundreds" of dollars per square foot. As the reporter paraphrased Belomy, it was a building for computers, not people.

As I wrote in Bandwidth, the pull outlasted the companies that created it. When WorldCom emerged from bankruptcy in 2004 as MCI, it moved its corporate headquarters to Ashburn. We'll come back to Loudoun County in Chapter VI. By then the farmland would be known by a different name.

The Build-Out—Project DAN, Modems and the Trash Can#

Behind the exchanges and the carrier hotels was a simpler story. The networks were buying hardware as fast as they could install it, and they were installing it in rooms.

UUNET

As I wrote in Bandwidth, augmenting capacity on the UUNET network was expensive. Dozens of servers sat in data centers across the country, connected by lines leased from fiber companies, many of them from MFS. UUNET's chief scientist put it best: "The project was swallowing cash by the trash-can full. We had to put hardware everywhere."

UUNET's appetite for bandwidth made it one of MFS's biggest customers. The Washington Post later reported that local phone fees made up about 40 percent of UUNET's costs, and MFS had fiber in the big cities. In February 1993, Crowe offered Adams $8 million for the company. "I wanted $12 million, and they wouldn't consider spending that much," Adams recalled in Bandwidth. "Everyone thought I was crazy to walk away, but I was having fun and I thought the Internet business was a good bet."

Venture money followed instead, and with it a professional manager, John Sidgmore, from GE. Sidgmore convinced Microsoft to take a 17 percent stake and use UUNET as the backbone for its online network. A month later he signed AOL. In May 1995 UUNET went public. The Post reported that the shares priced at $14 and closed the first day at $26. By November the stock hit $93.25, and UUNET was worth close to $3 billion.

Wall Street had found a new industry. In February 1996, Morgan Stanley's Mary Meeker and Chris DePuy published The Internet Report, a research report on the companies building the Internet that was later published as a book.Accounts differ on the date. Some put the report in 1995; I have used February 1996. I have not read a copy, so I don't quote from it here.

Project DAN

My own part of this story ran through the phone switch. In 1993, as I wrote in Bandwidth, Royce Holland handed me an assignment to open local voice competition. The technical name was cocarrier: a competitor and a Baby Bell would interconnect as equals. My team negotiated the nation's first cocarrier agreement, with Ameritech. In 1996 MFS became the first competitive local exchange carrier in the United States to receive approval from the FCC.

The voice business taught me to stay away from voice. But the same platform turned out to be perfect for the Internet. In the dial-up era, consumers reached the Internet through phone calls. At busy hours in the evening they would try over and over to log in. Once connected, they never logged out. Under the cocarrier rules, the Bell companies paid their rivals by the minute to complete those calls. Modem traffic drained the Bells' cash flow and enriched MFS.

UUNET had an initiative called Project DAN, for dial-up access network.In the dial-up era, the Internet was reached through modems attached to phone lines. A dial-up access network was the collection of equipment and data links that answered those modem calls and connected them to the Internet. Responsibility for executing it was assigned to my team, which led most people to believe the project was named after me. It wasn't. The work was network planning, network implementation and endless pestering of the Bell companies for more capacity between their networks and ours. We turned up capacity across the country and watched the revenue pour in.

That platform, as I wrote in Bandwidth, was why Sidgmore said yes to Crowe the second time. In late April 1996, MFS came back with $2 billion. The Wall Street Journal reported that MFS would double its capital spending to $1 billion a year, build its own long-haul fiber and reach eighty-five cities. Sidgmore explained the logic: "This is a scale game, and to prevail you have to get big fast."

Let's do the arithmetic. In February 1993, Crowe offered $8 million. Three years later he paid $2 billion. Adams had held out for $12 million and ended up with a stake worth far more.The sources disagree on Adams's payout. Bandwidth puts his take at $138 million. The Washington Post (Swisher, May 1996) reported that Adams, then thirty-nine, owned almost 15 percent, worth close to $300 million at the deal price, and that about forty UUNET employees had become millionaires. The difference may reflect shares sold or held before closing. Bandwidth's bibliography also dates the Swisher article May 5, 1996; ProQuest shows May 6.

Within months, the buyer was bought. In August 1996, as I wrote in Bandwidth, Crowe and Walter Scott sold MFS to Bernie Ebbers's WorldCom for $14.3 billion. A year earlier, Crowe had asked me to figure out why MFS's revenue was growing while its profits shrank. I told him. The sale was his answer.

AOL's Busy Signals

On December 1, 1996, AOL switched to a flat rate of $19.95 a month. Its customers had no reason to log off, and they didn't. Two and a half weeks later the Wall Street Journal reported the results. Peak load had tripled in a year to 230,000 simultaneous users. AOL added 8,400 modems and planned to spend $250 million on capacity by June. Pacific Telesis expected to spend up to $103 million handling Internet traffic, because online calls tied up lines about eighteen minutes longer than voice calls.

One frustrated subscriber, Raymond Toronto, gave the Journal the best description of the strategy: "This is like loading a million railroad cars full of merchandise and then deciding to build the tracks." AOL's Robert Pittman saw it differently: "People like to go to crowded places."

He was right, and the crowds needed somewhere to land. As I wrote in Bandwidth, UUNET's deal with AOL meant it carried most of the Internet traffic for the two largest Internet brands, and BBN signed a $55 million contract to provide backbone services to AOL that grew to $400 million a year.

The Modem Wave

WorldCom kept pouring money in. In February 1997 the Journal reported that WorldCom would invest $300 million in UUNET out of a $2 billion infrastructure budget. UUNET served more than fifty thousand businesses. People familiar with the plan said its modems would grow from 80,000 to more than 300,000. Sidgmore was blunt about the industry's capacity: "Everybody has been caught with their pants down."

By the end of 1997, WorldCom had paid $1.2 billion in stock and $175 million in cash for the networks behind AOL and CompuServe. The Post reported that the combined WorldCom, MFS and UUNET deal had been the fourth-largest merger in U.S. history.

PSINet, the third of the founding networks at MAE-East, ran the same race on less money. As I wrote in Bandwidth, Bill Schrader financed it at first with credit-card loans and the sale of a family car, and he took it public on May 1, 1995, at a valuation of $362 million. Schrader had a simple plan. "We're going to kill the phone companies, and we're going to be bigger than AT&T, and it will be a lot of fun," he told one executive.

Every one of these companies was building the same thing: rooms full of routers, modems and servers, wired to as many networks as possible. Most of them thought of those rooms as a cost of doing business. A few people were starting to think of them as the business.

The Building Is Worth More Than the Networks#

Look back at the numbers in this chapter. The federal government spent an estimated $12 million a year subsidizing the backbone it handed off in 1994. The NSF's award record for MFS Datanet shows $211,587 obligated. Crowe offered $8 million for UUNET in 1993 and paid $2 billion in 1996. WorldCom paid $14.3 billion for MFS a few months later.

Here is what I didn't understand at the time. At MFS, I thought our fiber was the asset. Fiber was hard to build and easy to price. But fiber could be overbuilt, and by the end of the decade it would be, many times over. The rooms where the fiber met could not. A network could be duplicated. A meeting place was valuable because everyone was already in it. Western Union's palace, the old Palo Alto phone office, a Tysons office building with an Ethernet switch: each one became more valuable with every tenant that moved in, and none of them could be copied by a competitor with a bigger budget.

The early operators gave away the most valuable part. MAE-East started as a free handshake. PAIX belonged to a computer company that sold it. The NAPs were run by phone companies with much bigger businesses to mind. Even MFS, which had built the most important exchange of all, thought of itself as a fiber company. We thought we were selling roads. The intersection turned out to be the prize.

Markoff's headline put the question in my own vocabulary. Will Commerce Flourish Where Rivers of Wire Converge? By 1997 the answer was obvious to anyone who walked into one of these buildings and listened to the generators.

The glass house had died on the desktop in the 1980s. In the 1990s, the network quietly rebuilt it, one basement, freight terminal and telegraph tower at a time. Only this time the glass house belonged to no single company. It was a place where everyone came to meet.

Somebody was about to start charging admission. A few years later, the Wall Street Journal would describe developers "scrambling to capture the bricks that house the clicks." The Gold Rush was coming, and it would start in the buildings.

In the summer of 2000, the hottest real estate around New York City was the real estate nobody had wanted a few years earlier. In Newark, the old Bamberger's department store had been converted for the Internet. In Manhattan, the Rudin family bought 32 Avenue of the Americas, the building where AT&T once ran its manual switching center, and filled it with MCI WorldCom and Qwest. Across the Hudson, a developer named Argent had bought six to seven million square feet of buildings in a single year. Exodus signed up for 230,000 square feet at 300 Boulevard East in Weehawken before Argent had even closed on the purchase.

Morgan Stanley's real-estate funds put $1.25 billion behind two developers of these buildings, Global Gateway and MetroNexis. In July 2000, the New York Times captured the mood in a headline: "Web Gives White Elephants a New Life."

Argent's executives explained the hurry. "Anything that will not be ready for 18 months is a non-starter," said Mark Teitelbaum. His partner Andrew Penson put eighteen months in perspective: "But that is like 10 years in Internet time."

Ten years in Internet time. That was how fast everyone believed the world was moving. And everyone wanted the same thing inside those old department stores and switching centers: racks of servers, powered and cooled around the clock, wired to as many networks as possible.

This chapter is the story of the companies that rushed to fill them. Exodus, born in a Sunnyvale room with no ceiling, became the hosting company of the boom. Two veterans of the Palo Alto Internet Exchange, Jay Adelson and Al Avery, founded Equinix on the opposite bet: a building where networks could meet without any one of them owning the room. Digex became the prize in a bidding war that Bernie Ebbers won and later regretted. The carriers piled in: Qwest with IBM, my own Level 3, PSINet, Genuity with its Black Rocket, and Gary Winnick's Global Crossing with GlobalCenter. Behind them ran a pack of hopefuls, including Globix, Data Return, NaviSite, AboveNet and Terremark. And out of MIT came Akamai, which bet that the Internet's future belonged at the edge rather than in a handful of giant buildings.

Every one of them was chasing a forecast. Forrester said Web hosting would grow 76 percent a year through 2003. IDC said Web services would go from $696 million in 1998 to $10.7 billion in 2002. Salomon Smith Barney saw hosting tripling to $9 billion by the end of 2001. Other analysts said $35 billion by 2005. No two forecasts agreed. All of them went up and to the right.

Exodus—K.B. Chandrasekhar, B.V. Jagadeesh, and Ellen Hancock#

The first Exodus data center was a single room in a Sunnyvale office suite, packed with about forty servers. The machines ran so hot that the founders had to knock out the ceiling and roll in portable air conditioners to keep them alive.

That room was the whole business. Two engineers from India, K.B. Chandrasekhar and B.V. Jagadeesh, had started out in 1992 with a network-software design firm called Fouress, incorporated in Maryland. By 1995, the company had been reincorporated in California as Exodus, and the founders had stumbled into a better idea: keeping other companies' web servers running, cooled and connected around the clock. Exodus began hosting servers in late 1995.

In August 1996, a $25 million private placement paid for the first real Exodus facility, in Santa Clara. Exodus called it an "Internet Data Center." Customers' servers sat in locked "cyber vaults" behind palm scanners.Palm scanners were the boom's favorite prop. They showed customers that someone was guarding the cash register. Chandrasekhar had a simple way of explaining why anyone would pay for this. "This is our customers' cash register," he told the Wall Street Journal. If a retailer's website went down, so did its sales.

Here is what strikes me about that line. In the 1960s and 1970s, companies paid Tymshare and EDS to run their computers for them. In the 1980s, the PC convinced everyone they would never need to do that again. Now the most aggressive new companies on earth were handing their servers back to someone else's building. In May 1999, the New York Times put the trend on its front page under the headline "Computing Centers Become the Keeper of Web's Future." A Goodyear executive gave the newspaper the best explanation I've seen of why: "We're in the tire business… Why run the digital power plant ourselves?"

The glass house was back. This time it had palm scanners.

The Seasoned Hand

Chandrasekhar and Jagadeesh were builders. Wall Street wanted an operator. In March 1998, days before the initial public offering, Exodus hired Ellen Hancock as president.

Hancock had a résumé that would have made anyone else in Silicon Valley nervous. She spent twenty-eight years at IBM and became its highest-ranking woman, overseeing fifteen thousand people. Lou Gerstner pushed her out. She moved to National Semiconductor as chief operating officer, then to Apple as chief technology officer in July 1996. A year later, Steve Jobs returned and ousted her too. When the Wall Street Journal profiled her in October 1999, the headline told the whole story: "Fired From Two Jobs, A High-Tech Manager Gets the Last Laugh."

She would get a lot more laughs before the last one.

The IPO

Goldman Sachs priced the Exodus offering late on March 18, 1998, at $15 a share. The company sold 4.5 million shares. The next morning the stock opened at $24.75 and closed the day up 84 percent.

The fundamentals were tiny. Exodus had booked $12.4 million of revenue in 1997 and lost $25.3 million. But investors weren't buying 1997. They were buying Forrester Research's forecast that Web hosting would grow from $200 million in 1997 to $16 billion by 2002. That's eighty times in five years. In the Gold Rush, the forecast was the asset.

The two men who had knocked out the ceiling owned surprisingly little of it. Going into the IPO, Chandrasekhar held 6.3 percent of Exodus and Jagadeesh 3.6 percent. Venture funds held most of the rest, led by the Fleet funds at 16.1 percent.

By late April, Hancock's stake was worth about $30 million on paper. Exodus already had sites in the San Francisco Bay Area, New York, Los Angeles, Seattle and Washington. When @Home took a $1.5 million stake to jointly market hosting services that same month, the Journal felt the need to explain the jargon to its readers: "co-location," or "co-lo in industry jargon." A new industry had a nickname.

The Rocket

Exodus spent 1999 doing what every Gold Rush company did. It built and it bought. It bought Cohesive Technology for $100 million in April. In October it paid $280 million in stock for Service Metrics, a website-performance company here in Boulder, Colorado.Service Metrics was backed by Mobius Venture Capital. In his foreword to Bandwidth, Brad Feld wrote that the Exodus sale "returned our entire first fund." By November, Exodus ran about a dozen data centers and expected to have twenty-two by the end of the year. It finished 1999 with nineteen, seventeen in the United States plus London and Tokyo, covering about 1.6 million square feet. A year earlier it had eight centers and 200,000 square feet.Exodus's 1999 10-K counts nineteen Internet Data Centers at year-end: four in Silicon Valley, thirteen more in eight other U.S. metros, London (opened June 1999) and Tokyo (acquired with Global OnLine Japan in December 1999). Its 1998 10-K counts eight U.S. centers, about 200,000 gross square feet and "over 800 customers." By the end of 1999 the company had "over 2,200 customers under contract" and revenue of $242.1 million, with a net loss of $130.3 million. Its Market Value in November was almost $7 billion. One analyst called it "an absolute rocket of a stock."

The Chorus

Wall Street had been cheering since spring. In late March 1999, David Levy of ING Baring Furman Selz started covering Exodus with a buy. The stock jumped 28¼ points to 150 that day, and InternetNews credited the note with sparking a rally across Internet stocks. One analyst's report had moved a whole sector.

By the end of the year, the cheering was unanimous. According to Zacks, fifteen analysts rated Exodus a strong buy and eight more rated it a buy. Nobody said hold. Nobody said sell. Thomas Weisel's James Linnehan raised his twelve-month target to $170 from $65. "Exodus has a dominant market position in what we believe is the engine of the Internet, Web hosting," he explained. By Media Metrix's count, Exodus hosted more than 40 percent of the top websites.

Twenty-three buys. Zero sells. As I wrote in Bandwidth, Jack Grubman became the highest-profile analyst of the fiber boom. Hosting didn't need a Grubman. It had a choir.

One voice was off-key. Joel Yaffe of Giga Information Group looked at the same stock and saw something else. "Right now, there's just a lot of speculation going on," he said.

Three months later, the rocket had tripled. In February 2000, the Wall Street Journal reported that Exodus had a Market Value of more than $20 billion. Chandrasekhar had already moved up to chairman and was launching his next startup. His Exodus shares were worth more than $700 million, and he had become one of the leading "angels" for a generation of South Asian founders in Silicon Valley.

That's a long way from a room with no ceiling.

I never met Hancock or Chandrasekhar. At Level 3 we were racing into the same business. By March 2000 we had about 4 million square feet across thirty facilities, and I'm certain we chased some of the same customers. I can't recall a single one.

Switzerland

By mid-2000, hosting had a clear leader and a clear runner-up. Exodus was No. 1. No. 2 was GlobalCenter, the hosting unit of Gary Winnick's Global Crossing. GlobalCenter hosted about three hundred websites, including Yahoo and eToys. By the count Exodus later filed with the SEC, it had more than six hundred customers, among them a young search company called Google.Exodus's December 2000 merger proxy lists GlobalCenter's customers as including Google, Yahoo!, Akamai, eToys, the New York Times, the Washington Post and VISA USA. GlobalCenter lost $269.0 million from continuing operations in the first nine months of 2000.

In December 1999, Global Crossing had named Leo Hindery chairman and CEO of GlobalCenter, with a plan to take it public within six months. Hindery had just left AT&T's cable business, and he was careful about his non-compete. Hosting was neutral ground, he explained: "This is Switzerland."

Switzerland was for sale seven months later. The merger proxy Exodus filed that December tells the courtship almost day by day. On June 26, 2000, Thomas Weisel Partners told Hancock that GlobalCenter "might be interested" in a combination. A week later, GlobalCenter handed Exodus a draft merger agreement. Exodus's board told management to keep talking, "but to pursue a lower valuation."

In July 2000, the Wall Street Journal reported that Global Crossing was in talks to sell GlobalCenter to Exodus for $6 to $7 billion in Exodus stock. Then the news leaked. Exodus shares jumped 49 percent. Since Exodus was paying in shares, the deal's value jumped with them, to more than $8 billion. Global Crossing's board approved terms on July 14. Exodus came back with a lower price, Global Crossing "rejected the price proposal," and the talks stalled.

On August 1, Exodus's chief financial officer, Marshall Case, called Hindery to say Exodus would do the deal only on the terms it had offered on July 22. Hindery "rejected these terms" and said he intended "to pursue alternative transactions with other companies." He put it in a letter to Hancock the same day.

Hancock didn't wait around either. The proxy says only that Exodus spent August in talks with "another company." The dates match Digex, the hosting company controlled by Intermedia Communications, whose auction we'll see from the inside shortly. By August 20, Exodus had agreed to pay 1.85 of its shares for each Digex share. Global Crossing, meanwhile, pitched Intermedia a different deal: put Digex and GlobalCenter together. Intermedia balked at the strings attached to the Exodus stock, and on August 30, Exodus formally withdrew its offer.WorldCom's S-4 tells the Exodus–Digex talks from Intermedia's side. The parties agreed on 1.85 Exodus shares per Digex share between August 17 and 20. Intermedia's bond indentures required it to turn at least 85 percent of any Exodus stock into cash within 270 days, and it said it needed to raise $2.4 billion in that window. Exodus's "final offer," on August 25, was $250 million in cash plus 1.70 shares. Press reports put Exodus's bid at roughly $120 a share, about $8.5 billion; the filing gives only the exchange ratios, and the "$120 per share range" in it was Bear Stearns's guidance to WorldCom. The Exodus proxy does not name the "another company"; the match to Digex is my inference from the dates. That afternoon, WorldCom called. Within forty-eight hours, Bernie Ebbers had agreed to buy Intermedia, and with it Digex, for about $6 billion. The deal was announced on September 5, 2000. Recall from Bandwidth that Ebbers was steaming. According to Sidgmore, Winnick had called Ebbers to say Global Crossing was going to buy Digex, and Ebbers resolved to get it first. Sidgmore's diagnosis, as he told the Washington Post in 2002: "He was afraid he was going to look bad from a testosterone point of view."In December 2000, Delaware Chancellor William B. Chandler III let the WorldCom–Intermedia deal proceed but wrote that "this Court seriously questions the integrity of the Section 203 waiver decision." Section 203 is Delaware's anti-takeover law. The settlement gave Digex's minority holders about $165 million of WorldCom stock. Sidgmore was blunter two years later: "That is the stupidest deal of all time. We paid over $6 billion for a company that's worth maybe $50 million." (Goodman and Merle, Washington Post, June 30, 2002. Bandwidth dates the article June 29; ProQuest shows June 30, page A1.)

Hindery had walked out on August 1. On September 1, the day Digex slipped away to WorldCom, he picked up the phone. In the proxy's words, "Mr. Hindery contacted Ms. Hancock to renew discussions."

The two losers went back to each other. On September 28, three weeks after the Digex loss, Exodus agreed to buy GlobalCenter for about $6.1 billion of Exodus stock. At announcement, Global Crossing would own about 17 percent of Exodus.The Journal's 17 percent was the figure at announcement. Because the number of Exodus shares floated with a formula, the December 2000 proxy estimated Global Crossing's stake at "approximately 20.3%," or 108.2 million shares, at November 27, 2000, prices. Exodus also signed a ten-year commitment to buy network capacity from Global Crossing, which the Journal reported might be worth more than $4 billion.The proxy gives no dollar value for the commitment. It says Exodus "committed to purchase at least 50% of its network needs worldwide other than Asia" from Global Crossing. Global Crossing in turn agreed to use Exodus as its exclusive Web hoster, to stay out of hosting for two years and to grant Exodus a $100 million usage credit.

Together, the No. 1 and No. 2 hosting companies had thirty-two Internet Data Centers and almost four thousand customers. Hancock's 3.8 million Exodus shares were worth $187.6 million. Hindery, whose company was disappearing into hers, summed it up: "I'm a broken man, she gets to run it."

Two and a half years after the IPO, the executive fired by Gerstner and Jobs ran the combined No. 1 and No. 2. The deal closed on January 10, 2001.

Hancock had won the auction. She had also just tied Exodus for the next ten years to a company that, as I wrote in Bandwidth, would forever be remembered as "Global Crashing."

Equinix—Jay Adelson and Al Avery#

Jay Adelson and Al Avery had already built one neutral meeting place for the Internet. They quit their jobs to build thirty more.

We met their first one in Chapter II. Avery was the general manager of the Palo Alto Internet Exchange, or PAIX, the exchange that Digital Equipment Corporation opened for commercial service in July 1996. Adelson was its operations manager, and before that he had been a founding member of Netcom. The two of them designed, built and ran PAIX. In June 1998 they left Digital and incorporated a new company in Delaware. They called it Equinix.The Equinix prospectus says the company was "incorporated in Delaware in June 1998." In December 1999 the Wall Street Journal called it a "16-month-old company," which would put the start closer to August 1998. I've used the prospectus date. Within a year, the exchange they had left behind would be sold to AboveNet, which ran an exchange of its own, as we'll see later in this chapter.

Their idea was the mirror image of Exodus. Exodus wanted to run its customers' servers. Equinix wanted to run the room where the networks met.

Here is why that mattered. By 1999 the Internet had plenty of exchanges, the places where one network handed traffic to another. But the existing ones were owned by carriers: MCI WorldCom, SBC and AboveNet. If you ran a network, exchanging traffic inside a competitor's building meant your rival was your landlord. It knew your traffic, set your rent and decided how fast you could grow. Adelson and Avery proposed a building owned by someone who sold no network at all. Any carrier, Internet provider or big website could move in, and any tenant could connect to any other by running a cable across the floor.That cable is called a cross-connect: a physical link between two tenants' equipment inside the same building. It is the smallest product in the data-center business, and it will matter a great deal in later chapters. The more networks inside, the more valuable the building became to the next one.

The pitch fit in a sentence. As the Wall Street Journal reported in May 1999, "Equinix's biggest attraction will be its neutrality." Cisco, Microsoft, Stanford and Benchmark Capital put up $12 million at launch, the Journal reported.Equinix's prospectus shows only Benchmark and Cisco buying Series A shares in September 1998, at 67 cents a share (split-adjusted). Microsoft came in with the Series B in late 1999, at $5.33. The plan called for as many as thirty-five private exchanges.

The Race

Equinix signed its first customer in April 1999 and opened its first exchange in July, in Ashburn, Virginia, near Washington, D.C. Part of the build-out was financed with loans from Comdisco, the disaster-recovery company we met in Chapter I. Newark and San Jose were next. By November the tenants included MCI WorldCom, Concentric and Akamai, and the Journal was calling Equinix "the first independent manager of Internet data exchanges."

That same month, Equinix announced the real plan: about $1.2 billion over four years for thirty centers, built by Bechtel. A few weeks later, the company lined up $80 million of new venture capital and $200 million of senior notes.The $200 million of senior notes were due in 2007. Salomon Smith Barney led the deal, with Morgan Stanley and Goldman Sachs. AOL, Marc Andreessen, Pacific Century and Reuters joined as investors. The Journal called Equinix a "16-month-old company." It had raised $315 million in total.

Do the math. A $1.2 billion plan, $315 million in the bank and almost no revenue. Equinix didn't start recognizing revenue until November 1999, and its revenue for the whole year came to $37,000.The prospectus contradicts itself on the start date. Its summary says "we began recognizing revenues in November 1999"; its business section says December. Where would the rest of the money come from? In 1999, nobody asked that question twice. The answer was always the same: the public markets.

Adelson was not worried about money. He was worried about the clock. When the Journal wrote about Bechtel's dot-com business in March 2000, under the headline "Venerable Builder Bechtel Gets a New Lift From the Dot-Com Boom," Adelson explained why Equinix had hired the company. "Our biggest competitor is time—being the first to market with this product—and that is exactly what Bechtel gives us," he said.

Bechtel had taken stock warrants in its deals with Equinix, Webvan and iMotors. When a builder the Journal called "venerable" starts taking warrants from Webvan, you know where you are in the cycle.

Grown-Ups and an IPO

Like Exodus, Equinix brought in a seasoned operator before going public. In May 2000, Peter Van Camp became CEO. He came from UUNET, where he had been president of Internet markets and the Americas. Avery moved over to president and chief operating officer. Adelson kept the title the Journal used for him: founder and chief technology officer. The prospectus gave the buildings a name of their own: "neutral Internet Business Exchange, or IBX centers."In June 2000, Equinix also assigned its purchase of eighty acres of land in San Jose to iStar. Its customers ranged from AT&T, WorldCom and my own Level 3 to Enron. And it was candid about the bet at the heart of the company: "the neutrality aspect of our business model is unique and largely unproven."

The timing was not kind. The Nasdaq had peaked at 5048.62 on March 10, 2000. Equinix went public five months later, on August 11, 2000. Goldman Sachs sold 20 million shares at $12, the top of the $10-to-$12 range. The stock rose 12 percent on its first day, to $13.38. Benchmark Capital, which had bought its first shares at 67 cents, went into the offering owning 16.1 percent of the company.

Twelve percent. Exodus had jumped 84 percent on its first day two years earlier. Akamai, as we'll see shortly, jumped 458 percent. In August 2000, twelve percent counted as a good day.

Equinix had raised its money and broken ground just as the music was slowing. Whether it could finish building before the music stopped would define its next three years.

Digex—Bernie Ebbers#

Digex was the hosting company that everyone wanted and almost nobody could buy.

We saw the Digex auction from Exodus's side of the table. Here it is from the other side. Digex was based in Beltsville, Maryland, and its public shares traded on their own. But Intermedia Communications owned 54 percent of its equity and 92.5 percent of its voting control. Intermedia was a debt-laden regional phone company in the Southeast with ninety thousand customers in fourteen cities. Earlier in 2000, it had cut its revenue guidance and put itself in play.

The arithmetic was strange. In September 2000, Digex had a Market Value of $5.3 billion. Intermedia, its parent, was worth $1.24 billion. Intermedia's Digex shares alone were worth $3.3 billion. The tail was worth almost three times the dog. Anyone who bought the dog got the tail.

The dog came with a catch: more than $3 billion of debt that would come due if anyone bought the company.

Bear Stearns ran the auction for Intermedia. It contacted twenty-four companies, and ten signed confidentiality agreements.The count comes from Intermedia's account in WorldCom's S-4. The Delaware opinion says six parties signed confidentiality agreements. The court also has Miller's call coming between 6 and 7 p.m. on August 30, offering "$120 a share or more." Three serious bidders emerged. Global Crossing offered $22 a share for Intermedia, and it offered to contribute GlobalCenter, which it valued at $8 billion, to a three-way combination. Exodus offered its own stock for Digex, then walked away on August 30, when Intermedia's bond covenants and the limits on reselling Exodus shares couldn't be reconciled.

That afternoon, a representative of WorldCom called. Bear Stearns told him any offer for Digex "would have to be in a $120 per share range." WorldCom's banker, Scott Miller of Salomon Smith Barney, didn't flinch. According to the Delaware Chancery Court, he said WorldCom "would outbid anyone for Digex." Global Crossing, which had a deal nearly done, said it "would not compete with such an offer" and set a deadline of 5 p.m. on September 1. Late on August 31, Ebbers changed the target. Rather than buy Digex, WorldCom would buy all of Intermedia and leave Digex public.

The deal was signed at about 5 p.m. on September 1. As the Delaware court later put it, "Almost exactly forty-eight hours would elapse" from WorldCom's first call to approval.

On September 5, 2000, WorldCom announced that it would buy Intermedia for $39 a share in stock, close to $3 billion, plus the debt. "The total cost of the transaction could be double that," the Journal noted. It was WorldCom's first big deal since its $115 billion Sprint merger fell apart.

The market didn't like it. WorldCom shares fell 9 percent.The Wall Street Journal reported on September 6, 2000, that WorldCom shares fell 9 percent. In Bandwidth I wrote that WorldCom's stock dropped 17 percent when the deal was announced. I've used the Journal's same-day figure; the difference should be checked against the trading record before publication. Digex shares fell almost 20 percent, to $67.88. Digex's minority holders had just watched Exodus's bid for their company disappear, a bid the press had valued at $8.5 billion. A PaineWebber analyst said Digex fell "in part because of the way Digex minority shareholders were treated."

The Waiver

How Digex's board handled the deal invited a lawsuit. Five of its eight directors had ties to Intermedia. A special committee set up to protect the minority holders advised against the deal. The board granted WorldCom a waiver from Delaware's anti-takeover law anyway. The vote, on September 1, was four to three. Every yes came from an Intermedia director or officer. The special committee's members and Digex's own chief executive voted no. "Minority shareholders aren't getting much respect these days," the Journal observed.

TCW Technology LP, which owned about 4.3 percent of Digex, sued on September 20. Delaware's Chancellor William B. Chandler III found the claims "sufficiently viable" to fast-track. In December he let the deal proceed, but as noted earlier, he questioned the integrity of the waiver and warned that damages could reach $2.5 billion. By then WorldCom's stock had fallen far enough that the whole deal, debt included, was worth about $4.2 billion.

It was WorldCom's second legal setback in a year. "They used to do these things just rapid-fire," a CIBC analyst said.

Ebbers settled. In February 2001, WorldCom agreed to give Digex's minority holders about $165 million in WorldCom stock and Digex itself another $15 million. Intermedia's shareholders took a haircut, with the exchange ratio cut from as much as 1.1872 WorldCom shares to one. When Chandler approved it on April 6, 2001, the New York Times called it a "$420 million settlement" on a "$4.1 billion purchase." Along with the stock, it included up to $250 million in concessions and $900 million of loans for Digex.

What did Ebbers want with Digex? He planned to fold it into WorldCom. He had won the auction. As Sidgmore would later make clear, winning was the problem.

The Carriers—Joe Nacchio, Jim Crowe, Bill Schrader, and Paul Gudonis#

If hosting was where the money was going, the carriers figured it should go through their buildings.

Their logic was simple. Every hosted server needed bandwidth, and the carriers owned the bandwidth. Why let Exodus resell your fiber at a markup when you could build the data center yourself and sell the customer both? MCI WorldCom planned to spend $1.2 billion in 2000 alone on facilities in fifty markets. Sprint planned eleven data centers by the end of 2001 and eighteen by the end of 2002, in a market projected to pass $20 billion by 2005.

IBM

The oldest name in computing was in the hosting business too. In May 1999, the New York Times opened its front-page story on hosting inside IBM's server farm in Schaumburg, Illinois, which ran the websites of about four hundred companies. Hosting, the Times reported, was "a billion-dollar business, doubling annually." AOL was spending $520 million on its third server farm. IBM, AT&T, Exodus and Verio were the names IDC listed in its forecast.

IBM's Walt Grom described his company's new role with a straight face: "We're happy to do the plumbing."

Think about that for a moment. IBM, the company that built the glass house in Chapter I, had watched the PC empty it out. Now it was renting the glass house back to the companies that had left. When Qwest decided it needed data centers in a hurry, it called the plumber.

Qwest

Joe Nacchio had wanted hosting from the start. As I wrote in Bandwidth, his vision for Qwest in 1997 was a full-stack phone company that rivaled AT&T, including voice, data storage, consulting and website hosting.

In March 2000, Qwest made it real with IBM. IBM would design and build twenty-eight U.S. Internet data centers for Qwest within three years, on top of four that Qwest was already building. About $2.5 billion would flow each way between the two companies over seven years. Each center would cost about $90 million to build and run. IBM already managed about 133 data centers. Analysts projected hosting would be a $35 billion business by 2005.

"We have been in a demand curve that has basically been insatiable," said Qwest's Lew Wilks.

Qwest called the buildings CyberCenters. By November 2000, it was leasing 500,000 square feet at 555 Tuckahoe Road in Yonkers, a former Western Electric and Saks building, and putting more than $100 million of equipment inside. "We are in 11 sites now and we expect to do three more by the end of the year," said Qwest's real-estate vice president, John McGuire. Qwest's marketing director had a forecast ready: "Projections are that Web hosting will be an $11 billion business by 2002."

Notice the number. In March, analysts had said $35 billion by 2005. In November, Qwest's own marketer said $11 billion by 2002. Nobody stopped to ask why the forecasts were so far apart.

Level 3

I was in the middle of this one.

As I wrote in Bandwidth, the years from 1997 through 2000 are a thick Gold Rush fog for me. We were building fiber networks across the United States and Europe, metro networks in most of the big cities and a tier-one Internet Backbone. We also built data center facilities in most of the big cities. We hired Kiewit as our general contractor. The initial estimate for the nationwide build was $3 billion. We spent in the vicinity of $5 billion. As I mentioned in the Exodus story, the Wall Street Journal reported in March 2000 that Level 3, "since its founding two years ago... has gone from nothing to four million square feet in 30 facilities."

The same Journal article contains a quote I find remarkable. Jimmy Byrd, Level 3's senior vice president for global Internet protocols and gateway services, held a title that covered those buildings. In the middle of the boom, he warned, "There are too many [data centers] coming online at the same time."

In the same article, John Sidgmore of UUNET said the opposite: "All of us combined can't fulfill the demand."

Both men were right. They were just looking at different calendars.

PSINet

Bill Schrader never met a growth market he didn't want to own. As I wrote in Bandwidth, Schrader committed $100 million for the naming rights to the Baltimore Ravens' new stadium during the Gold Rush, and PSINet's enterprise value soared to $10 billion. By December 2000, it had fourteen hosting facilities.

Schrader sold hosting as part of the package. In February 2000, he said, "We have exactly what the largest telephone companies want—network, hosting, expertise, and customer base—but at this point, we're not for sale." When he bought Metamor Worldwide for $1.9 billion a month later, he explained, "What we had before this deal was the best long-haul fiber carrier and also the best [data center] facilities." He also launched a $1 billion program to take equity from dot-com startups instead of cash for PSINet's services. What could possibly go wrong?

Genuity

Genuity started life as a hosting company. Late in 1997, GTE bought a small San Francisco web host called Genuity, whose biggest customers included Microsoft and Excite. In 2000, GTE stuck the name on a much bigger business that combined BBN, its Qwest fibers and the original Genuity.The two Genuitys were connected in another way. Sam Mohamad, the vice president of sales at Exodus when it went public in 1998, came from the original Genuity.

Genuity went public in June 2000. Verizon, which was absorbing GTE's phone business, kept about 10 percent with the right to convert to about 80 percent later. The offering priced at $11, below its range, and raised about $1.91 billion. The stock opened at $10.75 and closed its first day at $9.4063. The Journal pointed to the losses: Genuity had lost $647 million in 1999 on revenue of $706.5 million, and AOL alone accounted for 52 percent of that revenue.

In September 2000, Genuity unveiled Black Rocket, a bundle of managed hosting, Internet access, security and e-business tools, backed by a $20 million marketing campaign and a tagline asking, "Do You Want to Change the World?" CEO Paul Gudonis declared, "We have created an entire new industry category: Network Services Platforms." He added, "And now here comes Genuity with a complete solution."

As I wrote in Bandwidth, many of us thought it was nonsensical.

Global Crossing

We already met Global Crossing's hosting play in the Exodus story. GlobalCenter, run by Leo Hindery, hosted about three hundred websites, including Yahoo and eToys, and was the No. 2 hosting company before it disappeared into Exodus. Global Crossing also signed on as the anchor tenant of a big new building in Miami, which we'll come to shortly.

Broadwing bought the web hosting company SmartNAP and built its own data center and hosting business. Sprint, Dell and even Intel tried their hands at it.Dell formed a unit to market to Web hosts in February 2000 and later ran its own hosting business, DellHost. Intel shut its Web-hosting operation in June 2002.

Every carrier wanted to be in hosting. The question nobody asked was whether there were enough customers for all of them.

The Rest of the Field—Globix, Data Return, NaviSite, AboveNet, and Terremark#

Behind the leaders and the carriers ran a crowd of smaller hosting companies. Each had its own angle. Each raised money on the same forecasts.

Globix

Globix had a Market Value of almost $1 billion in 1999. It was building a data center at 415 Greenwich Street in Manhattan, and as late as June 2001 it said its centers were 95 percent full.The 415 Greenwich Street detail comes from an April 2001 New York Times article by Jayson Blair, whom the Times later found to have fabricated and plagiarized stories. Globix's own 10-K confirms that it bought the building in September 2000 and opened it to customers in June 2001. Keep that number in mind. Full buildings, it turned out, were not the same as a healthy company.

Data Return

Data Return was founded in August 1997 by three former Microsoft engineers. It was backed by Compaq and by Level 3.

Data Return went public on October 28, 1999, at $13 a share. It raised $81.2 million for 18 percent of the company, which valued the whole thing at nearly $450 million. A few days later, the Journal reported that Microsoft had invested $5 million in the company too.

NaviSite

NaviSite was the hosting company controlled by CMGI. In June 1999, Dell and Microsoft each bought minority stakes in it. By February 2001, NaviSite was cutting staff.

AboveNet

The name AboveNet will be familiar to Bandwidth readers, but in this story it means something different. The original AboveNet was a San Jose company that, as I described it in Bandwidth, sold higher-layer network services that reduced Internet congestion by steering traffic onto the Internet Backbones. It also ran one of the carrier-owned exchanges Equinix was competing against. It went public in December 1998, and its stock more than quadrupled in six months. Along the way, it bought PAIX from Digital's owner, Compaq. The exchange Adelson and Avery had built now belonged to a hosting company.

In June 1999, Metromedia Fiber Network, the fiber company I wrote about in Bandwidth, agreed to buy AboveNet for 1.175 MFN shares for each AboveNet share. The Wall Street Journal valued the deal at $1.76 billion. The New York Times said $1.55 billion.The two newspapers used different share prices for MFN to value the same exchange ratio. MFN's Howard Finkelstein explained, "We will be able to extend the reach of our 'carrier's carrier' business." As I wrote in Bandwidth, it was a departure from MFN's fiber strategy, and investors marked MFN's shares down 11 percent.

Years later, MFN would emerge from bankruptcy under the AboveNet name, and Zayo would buy it. In 1999, nobody could have imagined that.

Terremark

Terremark came at the boom from real estate. By March 2000, its T-Rex unit, run by Preminger, Mulroy and Manny Medina, was buying twelve properties for telecom tenants.

Then Medina went bigger. In June 2000, a consortium called InternetCoast proposed a network access point, or NAP, in Miami. Terremark and the owners of the Miami Heat were building a 700,000-square-foot, $100 million building to house it. Global Crossing and its hosting unit would take about 250,000 square feet.

Was Medina worried about filling the rest? "My concern regarding the potential use is zero," he said.

The building became the home of the NAP of the Americas. Its anchor tenant, Global Crossing, would be bankrupt within two years.

The Landlords—Argent, Max Capital, and the Telecom Hotels#

All of these companies needed buildings, and a new kind of developer showed up to supply them.

The trade distinguished two kinds of building. A telecom hotel, or telco hotel, was where networks put their switches and routers and connected to one another. A data center was where companies put their servers. Developers wanted to build both, and in March 2000 the Wall Street Journal summed up their new product in a headline: "Buildings Become Centers For Switches and Servers, But Some See a Glut." That was the article in which Level 3's own Jimmy Byrd worried about too many buildings coming online at once.

The developers didn't wait for tenants. In September 2000, the New York Times reported that a firm called Max Capital was building a 240,000-square-foot data center in Newburgh, New York, on speculation. The same article reported that DoubleClick had gone looking for 100,000 square feet at $20 a foot. In New Jersey, Global Gateway and MetroNexis had Morgan Stanley money behind them. In Miami, Terremark had its 700,000 square feet. In Yonkers, Qwest had its old Western Electric plant.

Here is the question nobody seemed to be asking. The developers built with borrowed and invested money. The hosting companies signed the leases. Their customers included Yahoo and eBay, and also dot-coms like eToys, whose ability to pay depended on investors' willingness to keep funding them. If the investors stopped, who would pay the rent?

Argent's Andrew Penson had said eighteen months was like ten years in Internet time. Eighteen months from that summer would land at the end of 2001.

Akamai—Tom Leighton, Danny Lewin, and Paul Sagan#

Everyone else in this chapter was building bigger buildings. Akamai built a smarter network.

Danny Lewin was born in Denver and raised in Jerusalem. He served as a captain in an elite unit of the Israeli army, studied at the Technion and worked for IBM in Haifa before landing at MIT, where Tom Leighton was his professor. In 1998, Lewin, Leighton and Jonathan Seelig founded Akamai.

Their insight was about distance. A popular website served from one building, however big, had to send every page across the Internet to every visitor. When a million people showed up at once, the site choked. Akamai's answer was to put copies of its customers' content on servers all over the world, close to the people asking for it. "The whole name of the game here is to push content to the edge of the Internet, very, very close to the end user," Akamai's David Goodtree told the New York Times in April 1999.

Akamai had competition. The same Times article sized it up against a rival called Sandpiper Networks. But Akamai had MIT's math and a seasoned chief executive.

That was a different bet from Exodus's. Exodus wanted every important server in its buildings. Akamai wanted copies of every important website in as many places as possible. As it happened, many of those places were other people's data centers. Akamai was one of Equinix's earliest tenants.The Journal listed Akamai among Equinix's tenants in November 1999. Akamai does not appear on the customer lists in Equinix's 2000 prospectus or its 2000 10-K.

Grown-Ups

Akamai brought in its grown-ups early. In 1999 it hired George Conrades as president and CEO. Bandwidth readers know him as the BBN chief who sold his company to GTE. I never met Conrades, but as I mentioned in Bandwidth, Akamai later became one of Zayo's largest customers. Paul Sagan, forty years old, came from Time Warner, where he had worked on NY1, Road Runner and the Pathfinder website. He joined as president and chief operating officer and paid $533,100 for nearly three million shares.

Sagan had a crisp view of how badly the Web measured itself. "You have 'hits,' which don't really mean anything," he told the Times. "So you were always grabbing at clouds."

The Fourth-Biggest Pop Ever

Akamai went public on October 29, 1999. The shares priced at $26, opened at $110 and closed at $145.1875. The 458 percent first-day gain was the fourth-largest in history. At the close, Akamai had a Market Value of more than $13 billion. Its sales over the previous nine months totaled $1.3 million. Its network ran on 1,475 servers.

Thirteen billion dollars, on $1.3 million of sales. That's ten thousand times nine months of revenue.

By December, Sagan's $533,100 was worth $617 million. Days before the IPO, the Journal had already run the headline that told the rest of the story: "Professor Stands to Reap Windfall in Akamai IPO."The New York Times article on Akamai's first trading day said the company sold 9 million shares "to raise more than $1.3 billion." That can't be right: 9 million shares at $26 comes to about $234 million. I've left the figure out.

Akamai had one of the best technology stories of the boom, and one of the most expensive stocks. Both would be tested soon enough.

Four Bets#

Step back from the names, and the hosting Gold Rush comes down to four bets about where the Internet's computers should live.

The first was Exodus's bet, shared by Digex, Globix, Data Return and NaviSite. Put the servers in our building, and we'll keep them powered, cooled, connected and watched. The customer bought peace of mind, and the hosting company bought the building, the generators, the air conditioners and the people. It was the glass house, rented by the rack.

The second was the carriers' bet. Qwest, Level 3, PSINet, Genuity, Global Crossing, Sprint and MCI WorldCom all reasoned that the customer needed bandwidth anyway, so the data center should belong to whoever sold the bandwidth. Hosting was a way to sell more of the network.

The third was Equinix's bet. Don't run anyone's servers, and don't sell anyone's bandwidth. Own the room where the networks meet, and let them meet each other.

The fourth was Akamai's bet. Don't put the website in one building at all. Spread it across many buildings, close to the people reading it.

Each bet had a different exposure. A hosting company needed a constant flow of new customers to fill expensive buildings. A carrier needed its hosting business to pull traffic onto a network it had borrowed billions to build. A neutral exchange needed networks to show up, and they only showed up where other networks already were. And an edge network needed websites willing to pay for speed.

Which of these bets depended on dot-com startups that might not exist in two years? Which depended on carriers that might not exist either? And which one got more valuable every time a new tenant walked in the door?

The market didn't ask those questions in 2000. It priced all four bets as if they would all win.

The Grid—PG&E and Silicon Valley Power#

The first warning came from a direction nobody in this chapter was watching.

In February 2001, Qwest wanted to open a 45,000-square-foot data center in Sunnyvale, a few miles from the room where Exodus had knocked out its ceiling. Pacific Gas & Electric told Qwest it couldn't have the power until 2002. "They told us they couldn't get power until 2002," Qwest's Rick Weston told the Wall Street Journal.

Qwest wasn't alone. PG&E had requests from more than one hundred server farms, asking for 1,200 megawatts in total. California was in the middle of its electricity crisis, and statewide power use had risen 6 percent in 2000. In Santa Clara, the city-owned utility, Silicon Valley Power, had forty requests for large new loads. Eighty percent of them were server farms. If all of them were built, they would double the utility's peak demand of about 450 megawatts.

"They're using enormous amounts of power," said PG&E's Keith Reed.

The same story was playing out on the East Coast. One project proposed for the South Bronx would draw 180 megawatts at its peak.The 180-megawatt Bronx figure first appeared in Jayson Blair's April 2001 New York Times article and is confirmed by the Wall Street Journal. The same Blair article compared it with 87 megawatts for the World Trade Center; I've left that comparison out because no second source confirms it. According to the Yankee Group, U.S. server-farm space had grown from 3 million square feet in 1999 to 9 million in 2000, and was headed to 25 million.

This was the first time the data-center industry ran into the power grid. In 2001 the collision didn't last. Within months, as the next chapter shows, the demand behind most of those requests evaporated. Remember the numbers, though: one hundred server farms asking a utility for 1,200 megawatts, and a single building asking for 180. Two decades later, single AI campuses would ask for many times that, and utilities would once again tell data-center builders to wait.The AI-era power figures are covered in Chapter VIII. To take one example, the three gas plants Entergy is building for Meta's Richland Parish, Louisiana, campus will supply about 2,200 megawatts.

The Power Myth

The request lists made the problem look enormous. So did a forecast with a following. In May 1999, Peter Huber and Mark Mills published an article in Forbes headlined "Dig more coal—the PCs are coming." A widely cited claim put computers and networking at 13 percent of U.S. electricity.

On February 2, 2000, Mills took the argument to a House subcommittee hearing titled "Kyoto and the Internet." Every piece of information technology, he told the members, had "two plugs, one for bits, one for power." Joseph Romm of the Center for Energy and Climate Solutions testified at the same hearing and did not hold back. "It is rare that we know so conclusively that someone is as wrong as Mr. Mills," he said, and he put the error at "at least a factor of 8."

The scientists at Lawrence Berkeley National Laboratory went out and measured. Jon Koomey put computers and networking at about 3 percent of U.S. electricity, not 13. In July 2002, the lab published its findings on hosting. America's hosting data centers in 2000 covered about 9.5 million square feet, close to the Yankee Group's count, and their computer rooms averaged 50 watts per square foot or less. All of them together drew under 500 megawatts, about 0.12 percent of the nation's electricity. Five facilities with billing data all came in below 40 watts per square foot. Actual loads ran three to four times below what the buildings had been designed for.

Compare that with the request lists. PG&E alone had been asked for 1,200 megawatts. Every hosting data center in the country was drawing less than half of that.

The lab's Jennifer Mitchell-Jackson explained the gap. "Rated power is typically several times greater than actual power use," she said. "Unfortunately, it is often this maximum power density that is cited in the media." Her verdict: "Many reports of data center energy use are exaggerations."

The first boom's power forecast was the data-center version of the traffic myth: a real trend, projected far past the evidence. Mills would testify to Congress again in April 2025, forecasting another surge in power demand. That time, as we'll see, the meters moved too.

Before Midnight#

At the peak of the Gold Rush, the hosting companies were worth fortunes.

Exodus topped $20 billion. Akamai was worth more than $13 billion on its first day of trading. Global Crossing valued GlobalCenter at $8 billion. Digex was worth $5.3 billion, and the press valued Exodus's bid for it at $8.5 billion. Ebbers paid about $6 billion for the parent company to get it. Genuity raised $1.91 billion in its IPO. MFN paid $1.76 billion for AboveNet. Globix was worth almost $1 billion. Data Return, a two-year-old company started by three engineers, was valued at nearly $450 million.

The founders and executives did fine on paper too. Chandrasekhar's Exodus shares were worth more than $700 million. Hancock's were worth $187.6 million. Sagan's $533,100 had turned into $617 million.

Underneath the Market Values sat the debt. At the end of 1999, Exodus owed about $1.53 billion in senior and convertible notes. Its stockholders' equity was $17.6 million. The company that hosted more than 40 percent of the top websites was built almost entirely on borrowed money.

In Bandwidth, I tallied the fiber companies the same way. AT&T, Qwest and WorldCom each topped out near $100 billion, and Global Crossing and Level 3 each passed $50 billion. Soon the combined Market Value of all those companies would tank to $15 billion. The hosting companies were about to take the same ride, and they were about to take it faster.

Eighteen months after the peak, Exodus would be trading for less than a dollar a share.

On March 5, 2001, the Wall Street Journal marked an anniversary nobody wanted to celebrate. A year earlier, on March 10, 2000, the Nasdaq had closed at 5048.62. Now it stood at 2117.63, down 58 percent. Of the stocks on the Nasdaq, 428 had fallen more than 90 percent.

For the hosting companies, the reckoning came a few months later. In June 2001, the Journal ran a short article under the headline "Overcapacity Jitters Hurt Web-Hosting Stocks." Stephen Murphy, an analyst at CIBC, estimated that only 22 percent of the data-center space in North America would be used or sold that year. Gartner cut its forecast for growth in colocation demand from 25 to 35 percent a year to 2 to 5 percent.

Gartner's Ted Chamberlin summed it up: "There's a glut of data-center space." PSINet and a company called Colo.com were already in bankruptcy.

Think about what that 22 percent meant. Four months earlier, the Yankee Group had counted 9 million square feet of server-farm space in the United States, headed for 25 million. CIBC was now saying that most of the space would go unused or unsold that year.

The developers felt it too. In Boston, a telecom hotel was being converted into space for biotech companies. Wiring a building for telecom tenants cost $500 to $1,000 a square foot, and now that money was buried in walls nobody needed. Bradley Olson of Cornell compared the overbuilding to "four gas stations on the same intersection." In Jersey City, a $100 million conversion of 1.1 million square feet at 50 Dey Street was still going ahead, and MetroNexis and Global Gateway, the developers Morgan Stanley had backed, had eleven telecom hotels with about 6 million square feet between them. "The demand is still there, but it's thinner," Andrew Zezas of Insignia/ESG told the New York Times in May 2001. By October, Grubb & Ellis put vacancy in telecom real estate above 30 percent in most markets.

Then the forecasts gave up. When Dell sold its Web-hosting business to Sprint in August 2001, the Journal did the arithmetic in one sentence: "Two years ago, Web-site hosting was projected to be a $17 billion annual market by 2003. Today, analysts estimate... about $9 billion."

Recall Argent's line from the summer of 2000: eighteen months was ten years in Internet time. Eighteen months later, the hosting companies found out what ten years of bad news looked like.

The tide was going out. Let's see who was swimming naked.

Exodus Continued—Ellen Hancock#

When we last left Ellen Hancock, she had just bought her biggest rival with richly valued stock. It took the market about five months to decide what that stock was worth.

In June 2001, the same week the Journal counted the empty space, Exodus shares hit $4.15. Their fifty-two-week high was $69.

The forecasts had been the asset. Now the forecasts were gone.

The departures came in a familiar order. In May, Exodus's chief financial officer, its president and its head of marketing left. In August, the company hired Goldman Sachs, Credit Suisse First Boston and Lazard as advisers, and three directors resigned. The stock was down 98 percent. Hancock still insisted "we're getting good solid business."

Hancock resigned as chairman and CEO in early September 2001. Exodus shares traded at 65 cents. The company carried about $2.9 billion of debt.Press reports at the time put Exodus's debt at about $3.5 billion. I have used the figure in Exodus's bankruptcy disclosure statement: "Indebtedness for borrowed money as of the Petition Date totaled approximately $2.9 billion." A J.P. Morgan analyst said what everyone on the Street believed: Exodus "must sell to a suitor or face bankruptcy."

There was no suitor.

Even then, the analysts split. On September 25, Andrew Schroepfer of Tier 1 Research, a former Goldman Sachs hosting analyst, predicted a prepackaged bankruptcy, with bondholders getting 20 to 25 cents on the dollar. He saw "no way they could support the number of data centers they have out there." Cary Robinson of U.S. Bancorp Piper Jaffray held the other side: "We do not currently believe Exodus will go bankrupt."

Exodus filed the next day, September 26, 2001, fifteen days after the attacks on the World Trade Center. It went into Chapter 11 in Wilmington, Delaware, with about $600 million of cash, owing about $2.9 billion across eight note issues.The disclosure statement lists five series of senior notes, originally $1.95 billion, and three issues of convertible subordinated notes with about $993 million outstanding. Press reports described six bond issues totaling $3.5 billion. On September 25, eWeek put Exodus's cash at $340 million. At midyear it ran forty-four data centers around the world. Its floor space had grown from 600,000 square feet in June 1999 to 5.6 million by the month it filed. Customers included Yahoo and eBay, and according to the Financial Times, Exodus hosted forty-six of the world's top 150 websites. The Internet still needed Exodus's buildings. It just didn't need Exodus's balance sheet.

Hancock's successor, William Krause, wrote the epitaph in the press release announcing the filing: "We sacrificed profitability in exchange for growth and market share, over-expanding in some areas in advance of demand."

Two months later, Cable & Wireless agreed to buy the best of what was left. When the sale closed on February 1, 2002, it got twenty-eight of Exodus's data centers for $560 million in cash.The press numbers were all different. The Wall Street Journal reported that Cable & Wireless would buy "the best 30 of Exodus's 56 data centers" for about $570 million in cash plus $180 million of assumed liabilities; the Financial Times said "$850m cash." I have used Exodus's disclosure statement, filed in the bankruptcy case before Chief U.S. District Judge Sue L. Robinson. It says the buyer, Cable & Wireless's CWIS unit (formerly Digital Island), signed on November 29, 2001, acquired interests in twenty-eight data centers and "paid $560 million in cash upon closing," plus about $9 million for security deposits. The sale order was entered January 22, 2002, after a two-day hearing and about ninety-seven objections, mostly from customers, landlords and equipment lessors. Nobody else wanted them. In the words of Exodus's court filing, "No other parties, however, bid on the assets." The same filing was just as blunt about the rest: "The Debtors are not reorganizing their business; they are liquidating." The senior noteholders were expected to recover 14 to 24 cents on the dollar. The subordinated noteholders would get half a cent. The stockholders got nothing.

Let's do the arithmetic. In February 2000, Exodus had a Market Value of more than $20 billion. Twenty-one months later, its best data centers sold for less than 3 percent of that. Global Crossing's stake of about 20 percent, the payment for GlobalCenter, went down with the common stock. As I wrote in Bandwidth, the ten-year capacity commitment became basically worthless too.

The servers kept humming through it all. The buildings were fine. The buildings were always fine. And that is the lesson the next generation of data-center investors would take to the bank, as we'll see when Equinix and Digital Realty start shopping for empty floors at pennies on the dollar.

Digex Continued—Bernie Ebbers#

When we last left Bernie Ebbers, he had beaten Exodus and Global Crossing to Digex, then paid Digex's minority shareholders $165 million in stock to make their lawsuit go away. Part of the settlement was $900 million of loans from WorldCom to Digex. Ebbers had bought himself a hosting company that needed his money to keep building.

The money ran out at the top. As I told in Bandwidth, WorldCom filed for Chapter 11 on July 21, 2002. Eleven days later, the FBI arrested its chief financial officer, Scott Sullivan, and its controller on securities fraud and conspiracy charges.The full story of WorldCom's accounting fraud and its executives' trials is told in Bandwidth. Here it matters only because Digex went down with it. Ebbers's empire, built on acquisitions paid for in stock, came apart.

Digex was a footnote to the WorldCom collapse, but a telling one. In June 2002, a few weeks before the bankruptcy, Digex cut 7 percent of its workforce. That same month, John Sidgmore, WorldCom's vice chairman, gave the Washington Post his verdict on the deal I described in the last chapter. WorldCom had "paid over $6 billion," he said, "for a company that's worth maybe $50 million."

Do the math on that. In September 2000, Digex had a Market Value of $5.3 billion, and the press had valued Exodus's bid for it at $8.5 billion. Less than two years later, WorldCom's own vice chairman valued it at less than 1 percent of either number.

Regulators, meanwhile, forced WorldCom to sell most of Intermedia's other assets, and WorldCom had trouble finding a buyer.

Ebbers had won the auction. Winning cost him more than losing would have.

Globix#

When we last left Globix, it had a Market Value of almost $1 billion and a data center going up in Manhattan.

Globix said it didn't have an empty-building problem. In June 2001, while CIBC was predicting that most of the continent's data-center space would go unused, Globix said its centers were 95 percent full.

It had a debt problem instead. In January 2002, Globix announced that it would seek bankruptcy protection. The plan would turn about $600 million of debt into $120 million of new notes plus equity. By then, Globix's Market Value had fallen to $5.87 million. Globix filed in Delaware on March 1, 2002, with the plan already agreed. The bondholders took about 85 percent of the new stock. The common shareholders kept about 1 percent. Globix emerged on April 25, eight weeks after filing.

Almost $1 billion to under $6 million. And the buildings were 95 percent full.

That is the lesson of Globix, and it's the lesson of this chapter. A full building with too much debt is a bankrupt company. The building itself was fine.Globix's preferred holders received about 14 percent of the new stock. The plan was confirmed on April 8, 2002. Globix was also negotiating to give up part of its Globix House space in London.

Data Return#

When we last left Data Return, three former Microsoft engineers had taken their two-year-old company public at a valuation of nearly $450 million, with Level 3, Compaq and Microsoft among its backers.

In November 2001, a company called Divine agreed to buy Data Return for about $31 million in stock. Data Return's shares were trading at 72 cents.

From $450 million to $31 million in two years.The $450 million is the valuation implied by Data Return's October 1999 IPO price, not its peak trading value. The market had once valued Data Return at nearly fifteen times what a buyer would pay for it.

PSINet Continued—Bill Schrader#

When we last left Bill Schrader, PSINet's name was on a football stadium and he was telling anyone who asked that PSINet was "not for sale."

As I wrote in Bandwidth, Schrader stayed an optimist to the end. At PSINet's holiday party, he told employees, "AT&T is going to go down, but PSINet will survive."Bandwidth dates this party to December 2001. That can't be right: Schrader resigned in April 2001, and PSINet filed for bankruptcy in May 2001. The party was most likely in December 2000. His board was less sure. PSINet carried about $3 billion of debt, and one director later said Schrader "was more interested in building the infrastructure than in achieving profitability with what he already had."

In April 2001, PSINet reported a fourth-quarter loss of $3.2 billion and disclosed that it was in default on its lease obligations. The board asked for Schrader's resignation and got it. He lost most of his $1 billion in PSINet stock by refusing to sell. A PSINet insider described his trading discipline: "When the stock was at $30 he'd say 'I'm a seller at $75.' When it went to $60, he said 'I'm a seller at $100.'"

On May 31, 2001, PSINet filed for Chapter 11 in the U.S. Bankruptcy Court in Manhattan. The press release carried the dateline of its headquarters: Ashburn, Virginia. Its chief executive, Harry G. Hobbs, offered the diagnosis: "Our existing capital structure did not permit us to respond to the rapid changes in our markets."

By November, the Wall Street Journal reported that nobody expected a bid above $300 million for the company, against nearly $3 billion of debt. The headline put it in retail terms: "In Bankruptcy, PSINet Gets '90% Off' Tag." The Journal named Citic Pacific as the likely buyer. Jim Millstein of Lazard put it this way: "anything north of zero is taken very seriously."

North of zero turned out to be the right neighborhood. In February 2002, David Schaeffer's Cogent Communications agreed to buy PSINet's U.S. business for about $10 million, and it closed in April.The $10 million comprised $7 million for the business and $3 million the buyer had already paid for the right to do its due diligence. Cogent's own filing gives "a total cash purchase price of approximately $9.5 million" and dates the closing to April 2, 2002. The Wall Street Journal had named Citic Pacific as the likely buyer in November 2001; Cogent ended up with the U.S. business. Bandwidth describes the purchase as PSINet's assets; the Washington Post headline says its U.S. division. As I wrote in Bandwidth, Schaeffer bought several other fiber properties at distressed prices and built them into a public company worth multiple billions. Nice job, Dave!

Genuity Continued—Paul Gudonis#

When we last left Paul Gudonis, he had launched Black Rocket and asked the world whether it wanted to change.

In 2001, Genuity went shopping in France. It bought a French web hosting company, Integra, for $114 million.Bandwidth describes Integra once as "a French web hosting company" and once as "a French fiber provider." The Wall Street Journal headline called the buyer a "net-services firm" and put the price at $113.7 million. Genuity's own announcement described Integra as a European provider of managed hosting and e-business services. The French target's chairman said his company had decided "to find a big brother." Within weeks, Genuity cut its revenue outlook for the year and reduced its workforce by 12 percent.

Genuity's own big brother was supposed to be Verizon, which held an option to buy it. In July 2002, with three years left on the option, Verizon gave it up. Verizon wanted investors to know it would not be absorbing Genuity's $3 billion of debt or its $4 billion of 2001 losses. The forfeiture triggered a covenant on Genuity's debt.

Two days after Verizon's announcement, Forbes ran a story titled "Genuity Is Left on the Launching Pad." As I quoted in Bandwidth, Forbes reminded readers of the rocket and the tagline, then added: "But the '90s, alas, had just ended, and the world had already changed—and not for the better, if you were a tech investor."

Genuity's stock, already down to $2.59, fell to 29 cents. A day before Verizon's announcement, Genuity drew down $723 million on its credit line, with an unusual demand that the banks complete the transfers by noon. A spokesperson insisted the timing was "absolutely coincidental." Genuity filed for bankruptcy in November 2002.

The buyer was my old company. Level 3 agreed to pay $242 million for Genuity's assets, with adjustments in Level 3's favor. By the time the deal closed in June 2003, the price had fallen to $60 million in cash plus $77 million of prepayments to vendors. Genuity had raised $1.91 billion in its IPO three years earlier. Level 3 bought the whole thing, BBN included, for less than 5 percent of its peak valuation.

The Carriers' Buildings—Qwest and Level 3#

When we last left Joe Nacchio, Qwest had hired IBM to build twenty-eight CyberCenters, had leased 500,000 square feet in Yonkers and was waiting on PG&E for power in Sunnyvale. Qwest's Lew Wilks had called the demand curve "insatiable."

By 2001, Gartner was forecasting colocation demand growth of 2 to 5 percent a year. What happened to the CyberCenters is a gap in my sources. What happened to Qwest is not. As I told in Bandwidth, Qwest's own president warned Nacchio repeatedly in late 2000 and early 2001 about the gap between his financial targets and what the business units were projecting. Nacchio didn't change his guidance. In June 2002, Phil Anschutz flew to New York and told Nacchio he was being replaced, effective immediately. A month later, the U.S. attorney in Denver opened a criminal investigation. In December 2005, Nacchio was charged with forty-two counts of insider trading over his sale of $100.8 million of Qwest stock in 2001.

When we last left Level 3, we had gone from nothing to 4 million square feet in thirty facilities in two years, and our own Jimmy Byrd had warned that too many data centers were coming online at the same time.

Byrd was right. I'd like to tell you we all saw it coming. The record shows that one of us did, out loud, in the Wall Street Journal.

Global Crossing Continued—Gary Winnick and Leo Hindery#

When we last left Leo Hindery, he was "a broken man" handing GlobalCenter to Ellen Hancock in exchange for about 17 percent of Exodus.

Hindery didn't stay to watch. He resigned as Global Crossing's CEO on October 12, 2000, two weeks after the Exodus deal was signed. By the time the deal closed in January 2001, Global Crossing's stake, by then estimated at about 20 percent of Exodus, was worth $1.8 billion. By the end of 2001, as we saw, it was basically worthless.

Global Crossing followed its stake down. As I told in Bandwidth, the company filed for Chapter 11 in January 2002. Singapore Technologies Telemedia eventually agreed to buy 61.5 percent of the reorganized company for $250 million, which valued all of Global Crossing at $600 million. A couple of years earlier, it had been worth $50 billion. Gary Winnick, meanwhile, had managed to extract $734 million for himself.

Global Crossing had been the anchor tenant of Terremark's new Miami building, taking about 250,000 of its 700,000 square feet.

AboveNet Continued—Steve Garofalo and Bill LaPerch#

When we last left Metromedia Fiber Network, it had paid $1.76 billion for the original AboveNet, a hosting and exchange company in San Jose, and announced a vast expansion into Europe.

Bandwidth readers know how this one ends. But the data-center part of the story deserves its own telling here.

In 2001, MFN faced an existential problem. Nearly every distressed fiber provider was its customer, and those customers made up most of its $2 billion revenue backlog. As they failed, MFN's funding failed with them. A Citicorp credit commitment turned out to be conditional. A class action was filed against Steve Garofalo, the CEO who had founded the company. The stock fell below $1 a share. In April 2002, MFN's auditor, KPMG, said it couldn't review the company's quarterly numbers because its internal controls were so poor. That same month, Verizon and Genuity terminated their fiber agreements with MFN.

MFN filed for bankruptcy in May 2002. It emerged in September 2003 under a new name: AboveNet, borrowed from the San Jose hosting company it had bought four years earlier. Its estimated enterprise value was $230 million to $330 million. At the peak of the boom, it had been worth $10 billion.

Selling the Floors

The bankruptcy turned MFN's data centers into cash. While it was still in Chapter 11, the company sold real estate in Northern Virginia's data-center market for a combined $63 million, a steep discount to what the properties were worth.

Then Bill LaPerch took over. LaPerch was a West Point graduate and former army captain who had come up through NYNEX and MCI as an operations executive. He joined MFN to run operations after WorldCom absorbed MCI, and he became CEO of the reorganized AboveNet, its fifth in three years. He sold multiple data-center assets. In December 2004, Equinix bought AboveNet's data center in San Jose. In October 2006, Digital Realty bought more AboveNet centers.

LaPerch used the proceeds to build fiber to as many data centers as he could reach. He explained the strategy in a 2010 interview I quoted in Bandwidth: "we know that a significant amount of business is going to be focused around the data center infrastructure. So, we establish a fiber ring that hits the key carrier hotels and data centers." He no longer wanted to own the buildings. He wanted to connect them. As I wrote in Bandwidth, his ten-year execution of that strategy was extraordinarily successful. Zayo bought AboveNet in 2012.

But look at who was buying. The two companies that bought AboveNet's data centers, Equinix and Digital Realty, are the two central characters of the next chapter.

September 11—Danny Lewin#

On the morning of September 11, 2001, Danny Lewin was aboard American Airlines Flight 11. He was thirty-one years old.

Two years earlier, Akamai's IPO had produced the fourth-biggest first-day gain in history. Lewin had grown up in Jerusalem after being born in Denver, served as a captain in an elite Israeli military unit, studied at the Technion and worked for IBM in Haifa. At MIT, he and his professor, Tom Leighton, had worked out how to push the Web's content to the edge of the network. With Jonathan Seelig, they founded Akamai in 1998.

Lewin died aboard Flight 11.

The New York Times ran his obituary two days later under the plainest of headlines: "Daniel Lewin, Technology Executive, 31."

The attacks changed the data-center business in ways that took years to play out. The first change showed up within weeks. Banks wanted backup data centers. In October 2001, the Wall Street Journal reported that a vacant 200,000-square-foot building in New Jersey had drawn a bidding war. Ron Vidal, a senior vice president at Level 3, explained the new thinking in eight words: "We don't want it all in one building."

The hosting companies had spent the boom selling concentration: put everything in our building, and we'll keep it safe. After September 11, customers wanted their computers in more than one place. It was the logic Akamai had been selling all along.

Pennies on the Dollar#

By 2003, the hosting Gold Rush had turned into a clearance sale.

Cable & Wireless bought twenty-eight of Exodus's data centers for $560 million in cash, and it had already paid $340 million for Digital Island. Cogent paid about $10 million for PSINet's U.S. business. Divine paid $31 million for Data Return. Level 3 paid $137 million for all of Genuity. ST Telemedia paid $250 million for control of Global Crossing. AboveNet sold its Northern Virginia data centers for $63 million.

In April 2002, Tier 1 Research tried to count the wreckage. Hosting companies around the world operated more than eight hundred Internet data centers, with over 40 million gross square feet between them. Less than half of that space was even in operation. Of the space that was open, 31 percent was in use. A year earlier, CIBC had estimated that only 22 percent of North America's space would be used or sold. Tier 1's count was hardly more cheerful.

Even the buyers weren't sure what they had. When Cable & Wireless bought Exodus's centers, a Gartner analyst offered a backhanded endorsement: "We think Web hosting is a valuable business... but it's one that's better for the large telecom operators." Many of the large telecom operators, of course, were the ones going bankrupt.

The buildings told a different story. Customers didn't leave them. Switching hosting providers could cost a customer 20 percent of a year's hosting bill, the Journal reported. When Cable & Wireless took over Exodus's best centers, the customer list it inherited included Microsoft and Cisco. The racks, the generators, the cross-connects and the customers all stayed put. Only the names on the deeds changed.

A buyer with cash and patience could pick up those buildings for a fraction of what it cost to build them. Nobody put it better than the Equinix executive who explained his company's purchase of AboveNet's San Jose building: "We're paying pennies on the dollar for the San Jose building's assets. If we had to go out today and build a building from scratch like this, we would have to deploy $80 million in capital."

The starkest example surfaced in 2004. In 2000, Exodus had built twin 150,000-square-foot buildings at 2045 and 2055 Lafayette Street in Santa Clara. The property, a later prospectus said, "was built specifically for Exodus." Its new owners estimated that Exodus had invested "approximately $150 million in the electrical systems" alone. In May 2004, GI Partners, the fund that would soon take Digital Realty public, bought both buildings for $60 million.

The same patience was at work in Ashburn. Two developers, Lammot du Pont and Hossein Fateh, had bought land there in September 2000 and put up two data-center shells in 2001. A colocation provider leased them from mid-2001 through late 2003. Then one of them sat empty until 2005. In their own later account, du Pont and Fateh "began acquiring data centers that were distressed assets between 2000 and 2003." We'll meet them again.

What should we take from all this? In Bandwidth, the myth was that Internet traffic doubled every three months. Andrew Odlyzko debunked it in 1998. As he and K. G. Coffman later showed, the doubling had been real in 1995 and 1996. By 1997, traffic had slowed to roughly doubling once a year. The myth was a true number that outlived its time. The hosting business had its own version of that myth. In 1998, Forrester said hosting would grow eighty-fold in five years. In 1999, forecasters said $17 billion by 2003. In 2000, analysts said $35 billion by 2005. By August 2001, the estimate for 2003 was about $9 billion. And a month later, in the press release announcing its bankruptcy, Exodus was still quoting IDC's forecast that hosting demand "will reach $24.8 billion by 2004, up from $7 billion in 2001." The buildings went up on the forecasts. The bankruptcies came when the forecasts came down.

The demand was real. It just wasn't as big, or as fast, as the forecasts said. Keep that distinction in mind when we get to the AI forecasts at the end of this book.

By the end of 2003, the worst of the Bust was behind the data-center business. The empty floors were still empty, but the prices had finally caught up with reality. Now came the part of the cycle that rewards the patient: the time when buyers with cash walk through the wreckage and pick up hard assets for next to nada. The table was set, once again, for the phoenix.

I opened the Phoenix chapter of Bandwidth with the same six words. They earn a second tour here, because the data-center business ran the same play as the fiber business, on the same calendar, and sometimes with the same sellers.

In December 2004, Equinix filed a short Form 8-K with the Securities and Exchange Commission. Most 8-Ks are forgettable. This one recorded the end of an era in a single number.

The filing said that an Equinix subsidiary had leased "an existing data center located at 1735 Lundy Avenue, San Jose, California." The building held about 103,000 square feet. The lease ran "through 2020." Then came the line that mattered. Equinix had bought the data center infrastructure assets inside the building, for "approximately $900,000."The 8-K does not name AboveNet; Equinix's December 8, 2004 press release does. Equinix's 2004 10-K gives the price as $924,000. The "$80 million" quote that Bandwidth and Chapter IV attribute to the San Jose building matches the build cost Data Center Knowledge gave in June 2005 for a different Silicon Valley center. Both may be true.

Equinix's press release filled in the rest. The seller was AboveNet, the company that had come out of Metromedia Fiber's bankruptcy under Bill LaPerch. The building had room for roughly 2,000 cabinets. Equinix expected to spend about $5.0 million more to get it ready and to start placing customers there on March 1, 2005. At capacity, the company said, the center should bring in $25 million to $30 million a year.

Why would a seller take so little? As I wrote in Bandwidth, LaPerch sold multiple data-center assets in his first year as CEO, assets that, if held, could have been worth many hundreds of millions. He had a reason. He used the proceeds to push AboveNet's fiber into as many data centers as possible. LaPerch chose to own the roads into the buildings and let others own the buildings. As I wrote in Bandwidth, his ten-year execution of that strategy was extraordinarily successful. In 2012, my company, Zayo, bought AboveNet for about $2.2 billion.

Still, read those numbers again. Under $1 million for the guts of a finished data center, plus about $5 million of fix-up, against as much as $30 million a year of revenue. That was the trade on offer in the Phoenix years. Somebody else had paid to build the place. The new owner only had to fill it.

Equinix's next annual report turned the trade into a strategy. "In light of the availability of fully built-out data centers in select markets at costs significantly below the cost we would incur in building out new space," the company wrote, "we have altered our business strategy to acquire fully built-out data centers rather than build out our own." Stop building. Start shopping.

The fiber business was doing the same thing at the same time. In January 2002, as I quoted in Bandwidth, Forbes looked at the telecom wreckage and saw that "hard assets are selling for pennies on the dollar." I was one of the buyers who took that advice in fiber. In data centers, the buyers were a different cast, and this chapter is their story.

Equinix, a company trading for 30 cents a share in 2002, used a rescue from Singapore and a run of bargain buildings to become the place where the Internet's networks meet. A real-estate fund called GI Partners bought twenty-five technology buildings, many of them data centers, when nobody wanted them and turned most of them into Digital Realty. A young Las Vegas real-estate man named Rob Roy picked up a small building that Enron had wired for a bandwidth market that never came, and built Switch on top of it. Savvis, Terremark and Rackspace showed that the old Exodus business of running other people's servers could survive the Bust after all. And in 2008, a financial crisis bigger than the one that killed the hosting Gold Rush tested whether any of it would hold.

Equinix—Peter Van Camp and Steve Smith#

When we last left Equinix, Peter Van Camp had just taken it public at $12 a share, five months after the Nasdaq peaked, with a $1.2 billion building plan and nowhere near $1.2 billion in the bank.

The building plan did not survive the Bust. In early October 2002, Equinix stock closed at 30 cents. It was down 89.7 percent for the year. The company carried about $130 million of debt, and the bondholders owned a claim on a company whose shares traded like a lottery ticket.

From $12 to 30 cents in a little over two years. Exodus had made the same trip, and Exodus was in bankruptcy. Why didn't Equinix follow it?

The Singapore Rescue

The answer came from the far side of the Pacific. On October 2, 2002, Equinix signed a three-way combination with two other exchange companies. One was i-STT, the Internet-exchange arm of Singapore Technologies Telemedia. The other was Pihana Pacific, based in Honolulu.

Pihana was a Gold Rush company in miniature. It had raised $240 million in 2000 and built seven regional data centers, from Honolulu to Hong Kong and Sydney. By 2002, its Honolulu center was 40 to 50 percent occupied. Hong Kong and Sydney were about 30 percent full. Pihana had the buildings. It did not have the customers to pay for them.

The deal solved everyone's problem at once. Equinix took over Pihana's and i-STT's exchange centers in Asia and the United States. ST Telemedia agreed to invest $30 million in Equinix, and Equinix would use most of it to retire its bonds at a deep discount. When the dust settled, ST Telemedia would own 28 percent of the combined company, Equinix's existing shareholders 33 percent, Pihana's holders 21 percent and the senior bondholders 18 percent.The October 2002 agreements authorized "up to $40 million aggregate principal amount of convertible secured notes" for STT Communications and other purchasers. The investment at closing was $30 million. Equinix also cut its credit facility by $8.5 million, to $91.5 million, and Pihana brought more than $26 million of cash into the combined company. The bond retirement left about $30 million of the 13 percent notes outstanding. "We feel we've solved that problem," Pihana's CEO, Rich Kalbrener, told the Honolulu Star-Bulletin.

Bandwidth readers will recognize the investor. ST Telemedia was the same Singapore company that paid $250 million for control of Global Crossing out of bankruptcy, as we saw in the last chapter. In the Phoenix years, the cash sat in Asia, and Asia was shopping.

The combination closed on December 31, 2002. Stockholders had approved it the day before in New York. More than $116 million of Equinix's 13 percent senior notes, about 85 percent of them, were retired. The same day, a 1-for-32 reverse stock split took effect so that Equinix could keep its Nasdaq listing. Lee Theng Kiat, ST Telemedia's CEO, became chairman. Van Camp stayed on as CEO. Phil Koen moved to Singapore to run the Asia-Pacific business.

Van Camp was a UUNET man. Before Equinix, he had been UUNET's president of Internet markets and the Americas. Bandwidth readers know how UUNET's story ended, inside WorldCom. Van Camp's new company had come close to its own ending. The Singapore money bought it time.

Van Camp's statement that day was a survivor's boast. "Equinix is the largest network-neutral Internet exchange services company internationally," he said. The deal, he added, would "result in a strengthened balance sheet that provides Equinix with a solid financial base."

A company that had needed a reverse split to stay listed was calling itself the largest in the world. Was he right? In a narrow sense, yes. Many of its competitors were in bankruptcy court.

Buy, Don't Build

Equinix came out of the rescue small. In 2003, it booked $117.9 million of revenue and lost $84.2 million.The 2003 loss figure ($8.76 per share) comes from the selected-data table in Equinix's 2003 10-K, not from a narrative statement. But its balance sheet could now carry it through the time it needed, and the market was about to hand it the cheapest raw material in the history of the business.

In December 2003, Equinix signed a deal to sublease Sprint's E|Solutions Internet Center in Santa Clara. The center held 160,000 square feet. Equinix took over the building on a lease and bought the data-center property and equipment inside it. The Sprint center "became our 14th IBX hub," the company wrote. A carrier had built it during the boom. A neutral exchange would fill it during the bust.

A year later came the AboveNet building on Lundy Avenue, for about $900,000. Six months after that, in June 2005, Equinix took a fourth Silicon Valley center on a long-term lease. Data Center Knowledge reported the history: "The new center was built-out in 2001 at an estimated cost of $80 million, or better than $650 per square foot." Equinix planned to spend $15 million upgrading it.Data Center Knowledge did not name the seller of the fourth Silicon Valley center. The lease began in October 2005 and covered 120,000 square feet.

Here is the mechanic that made those buildings worth more in Equinix's hands than in anyone else's. Equinix did not sell bandwidth, and it did not run its customers' servers. It rented space in buildings where many networks met, and it charged for the connections between them. Every new network in an IBX made the building more valuable to the next customer, and every new customer made it more valuable to the next network. A carrier like Sprint competed with every other carrier in its building. AboveNet sold fiber. Equinix had nothing to sell but the meeting.

That was the bet Jay Adelson and Al Avery had made in 1998. It took the Bust, and a fire sale of other people's buildings, to prove it.

The Seasoned Hand, Again

In October 2006, Equinix announced that Van Camp would hand over the CEO job. On March 19, 2007, it named his successor: Stephen M. Smith, a former Hewlett-Packard executive. Smith took over on April 2, 2007. Van Camp, chairman and CEO since 2000, became executive chairman. "I'm confident Steve will serve Equinix customers and shareholders well," Van Camp said.

I should say up front that Steve Smith is a friend of mine. He ran Equinix for about a decade, and years later he succeeded me as CEO of Zayo. I'll try to tell his story the way I'd tell anyone's, from the record.

Smith's résumé tells you what Equinix's board was looking for. He served eight years in the U.S. Army. He then spent sixteen years at EDS, where he rose to chief sales officer and president of EDS Asia-Pacific. After that came Lucent, as a vice president, and then HP Services, as a senior vice president.

EDS. Readers of Chapter I will remember it. In the 1960s and 1970s, EDS was one of the companies that ran other people's computers in the original glass house. Now an EDS-trained salesman was taking over the company that owned the rooms where the Internet's networks met. The founders had built Equinix as an engineering idea. Smith had spent his career selling and running services.

He moved fast. In June 2007, less than three months into the job, Equinix agreed to buy IXEurope for $482 million.The price was 125 pence, or about $2.50, per share. IXEurope had fourteen data centers with 380,000 square feet in London, Paris, Frankfurt, Munich, Düsseldorf, Zurich and Geneva. "The deal provides Equinix with a European presence to go with its network of facilities in the U.S. and Asia," Data Center Knowledge wrote. After the deal, Equinix's footprint topped 3 million square feet.

Five years earlier, the company had needed Singapore money to stay alive. Now it had exchanges on three continents.

The Juggernaut

Equinix still had a hole in North America. Its buildings sat in the biggest metros. The second-tier cities belonged to others, and the largest of those others was Switch and Data.

The two companies had talked once before, in 2006. Those talks ended, Equinix's merger proxy later explained, "because the expected value of Switch and Data following the initial public offering significantly exceeded the value that Equinix had indicated it would be willing to offer." Switch and Data went public in early 2007 instead.

Smith tried again in the spring of 2009, with the markets still raw from Lehman. On May 4, he, his CFO, Keith Taylor, and Mark Adams sat down with Switch and Data's chairman, William Luby, at Luby's office in New York. Smith asked about a combination. Luby's answer, as the proxy records it: "Switch and Data was not for sale."

Fifteen days later, Smith broke off. Equinix's management, the filing says, "was focused on other priorities." On July 8, he called to resume talks. A week later he floated "a 20-30% premium," mostly in stock, and asked for exclusivity. Luby refused.

Luby had a reason to hold out. Since January, a private, non-U.S. company that the proxy calls only the "International Party" had been talking with Switch and Data about a stock-for-stock merger of equals. Smith had competition.

So the price climbed. Equinix's letter of intent on August 28 offered $16.35 a share, a 20 percent premium. On September 21, it went to $18.50. Switch and Data countered the next day at $19.47 and asked for a forty-five-day go-shop instead of exclusivity. On September 24, Equinix settled at about $18.46 a share, a premium of about 30 percent. The next day, Luby signed twenty-one days of exclusivity and told the International Party that the talks were over.Every offer was 80 percent stock and 20 percent cash. The steps, per Switch and Data share: 0.19276 Equinix shares or $16.35 (August 28); 0.20170 or $18.50 (September 21); Switch and Data's counter of 0.21000 or $19.47 (September 22); an Equinix offer of 0.20351 or $18.80 on a date the proxy doesn't give; and 0.20422 or $18.46 (September 24). The implied values moved with Equinix's share price, which is why a higher exchange ratio on September 24 shows a slightly lower dollar figure. The premiums are the proxy's own.

Then came the twist. While the lawyers drafted, Equinix's stock kept rising. By October 19, it had reached $98.39. Because most of the price was paid in Equinix shares, the agreed exchange ratio now implied $20.09 a share, a 39.3 percent premium. Smith cut the ratio. "Equinix's board of directors was unwilling to pay such a premium," the proxy says. The new terms were 0.19409 Equinix shares or $19.06 in cash for each Switch and Data share. Luby pushed back on how the cash portion would move with Equinix's stock price, but he didn't walk. Two days later, the boards signed.

That kind of re-trade only works when the buyer's currency is strong and the seller has already sent its other suitor home. Equinix's board, the proxy adds, chose "not to approach the other participants in the industry." Smith had one target, and he got it.

On October 21, 2009, Equinix announced that it would buy Switch and Data for about $689 million in cash and stock. Switch and Data operated thirty-four sites in the United States and Canada, in cities such as Atlanta, Denver, Miami, Seattle and Toronto. Smith said the deal "further strengthens Equinix's position as the most comprehensive global data center services provider."

The deal closed in the spring of 2010, one week after the Justice Department finished its antitrust review. By then, the value had slipped to $683.4 million: about $134.0 million in cash and about 5.5 million Equinix shares.Each Switch and Data share received either 0.19409 Equinix shares or $19.06 in cash, prorated to 80 percent stock and 20 percent cash; the proxy confirms these terms. The $683.4 million excludes assumed employee equity awards. Equinix's 8-K dates the closing to April 30, 2010; its completion release is dated May 3, 2010. Data Center Knowledge wrote that the deal closed "about a year" after its announcement; it closed about six months later. In January 2010, the same publication rounded the price to "$700 million." How big was Switch and Data next to Equinix? One research firm, Nortia, estimated that Switch and Data had about 8,920 cabinets billing and 5,080 more available, against about 23,600 billing cabinets at Equinix in the United States. In other words, Equinix was adding more than a third to its U.S. base in a single deal. Equinix went from six North American metros to twenty-two. The combined company had eighty-seven IBX and partner centers in thirty-five metros and more than 6 million square feet. Data Center Knowledge called the result "a colocation juggernaut spanning three continents."

Let's do the arithmetic. In October 2002, Equinix's stock traded at 30 cents and its bondholders were trading their claims for 18 percent of the company. Seven and a half years later, it handed over about $549 million worth of its own shares to buy a competitor with thirty-four sites.

The phoenix had a name, and the name was Equinix.

Digital Realty—GI Partners, Richard Magnuson and Mike Foust#

"Back then data center investing was an oxymoron," Richard Magnuson said in a 2010 keynote.

He was talking about 2002. Data centers were sitting empty or selling for very little. The industry was in a slump, and the space was oversupplied. The hosting companies that had built the buildings were in bankruptcy, and the carriers that had built the rest were trying to stay out of it.

Magnuson saw what the operators couldn't. The buildings were fine. The buildings were always fine.

The Fund

GI Partners, in the words of its later IPO prospectus, "was formed in February 2001 after a competitive six-month selection process." The selector was CalPERS, the California state pension fund, which made a $500 million equity commitment. Magnuson founded the firm to focus on data centers.The prospectus does not say who chose GI; the "selection process" was CalPERS's, which made the $500 million commitment. GI Partners later said CalPERS hired it after a request-for-proposal process in 2000. Data Center Knowledge's 2012 retrospective says Magnuson founded GI in 2001 to focus on data centers. The timing looked terrible. The Nasdaq had already crashed, and the hosting bankruptcies were just beginning.

That was the point. From 2002 to 2004, GI bought twenty-five technology buildings, many of them data centers, with $280 million of equity.The Digital Realty prospectus calls them "technology-related real estate properties," and several were not data centers: the Maxtor Manufacturing Facility, the ASM Lithography Facility and Ardenwood Corporate Park, for example. Its 2004 10-K says Digital Realty "acquired 24 technology-related real estate properties" between January 2002 and December 2004. The twenty-five figure and the $280 million of equity are GI's own, as reported by Data Center Knowledge. The first was the Univision Tower in Dallas, in January 2002. The Digital Realty prospectus explained the logic in one sentence: "the cost of acquisition is usually substantially lower than the cost to construct a new facility."

One purchase tells the whole story. In Santa Clara, at 2045 and 2055 Lafayette Street, stood two 150,000-square-foot buildings finished in 2000. "The property was built specifically for Exodus," the prospectus said. Digital Realty estimated that Exodus had put about $150 million into the electrical systems alone. In May 2004, GI bought both buildings for $60.0 million, fully leased to Savvis, the company that by then owned Exodus's remains. That was forty cents on the dollar for the wiring alone, with the buildings thrown in.

The tenants had their own troubles. Since GI's first purchase, the prospectus disclosed, fourteen tenants leasing about 474,000 square feet had gone through bankruptcy. Eight of them, with about 370,000 square feet, never missed a rent payment and kept their leases. When a company goes bankrupt, it stops paying for a lot of things. It keeps paying for the room where its servers live.

Here is how GI's model differed from Equinix's. Equinix was an operator. It ran the exchange, sold the cross-connects and staffed the building around the clock. GI was a landlord. It owned the building, the power and the cooling, and it leased the space to tenants who ran their own operations inside. The tenant signed a long lease. The landlord collected rent. When the tenants were big, creditworthy companies, the rent looked a lot like a bond.

That is a real-estate business, and Wall Street knows how to value real estate. GI would take it public as one.

The IPO

On October 28, 2004, Digital Realty Trust priced its initial public offering: 20 million shares at $12.00, or $240 million.The underwriters also had a 3,000,000-share overallotment option, of which 1,421,300 shares closed on November 30, 2004, adding about $15.9 million. The 2004 10-K puts the net proceeds from both sales at about $230.8 million, after $18.0 million of underwriting discounts and commissions and $8.3 million of other expenses. The deal closed on November 3, and the stock listed on the New York Stock Exchange under the ticker "DLR." The company was organized as a real estate investment trust, a structure that let it pass its rental income through to shareholders.

The portfolio held twenty-two properties in the United States and one in London, about 5.6 million net rentable square feet. At June 30, 2004, 87.1 percent of it was leased, on leases averaging 12.6 years. The largest tenant, at 12.6 percent of rent, was Savvis. Qwest was second. Magnuson became executive chairman. Michael F. Foust, a GI managing director, became CEO.

The IPO also put a price on GI's contrarian nerve. For the properties it contributed, GI received operating-partnership units worth $383.2 million at the offering price. The historical net tangible book value of what it contributed was about $114.6 million.

Twelve dollars a share. That was Equinix's IPO price four years earlier. Equinix had gone public on a forecast. Digital Realty went public on signed leases.

The Contrarian's Scorecard

GI had also bought from the Bust's survivors. In September 2004, it paid $36.5 million for a 179,489-square-foot building in downtown San Jose that had been redeveloped in 1998 for AboveNet's data center. Digital Realty estimated that AboveNet had invested "over $30 million" in improvements, and AboveNet stayed on as the main tenant. Three months later, the same seller let Equinix have the inside of a different San Jose building, on Lundy Avenue, for under $1 million.The two San Jose buildings were different facilities: the 179,489-square-foot downtown building that GI bought in September 2004, with AboveNet as the main tenant through November 2019, and the 103,000-square-foot center at 1735 Lundy Avenue whose lease and equipment Equinix took over in December 2004.

GI kept investing around the edges of the business too. Over the decade, it backed Savvis, The Planet, EV1 Servers, Telx, SoftLayer and ViaWest. GI Partners turned up on a lot of the survivors' cap tables.

In September 2009, GI sold its last data-center holding, a former Exodus center in Frankfurt. The buyer was Equinix, and the price was $28 million. Remember that Exodus had once been worth more than $20 billion. Its buildings were still changing hands, one at a time, among the companies that had figured out what they were worth.

By GI's own figures, the fund turned its $280 million of equity into more than $1 billion of proceeds, a 3.7x multiple and a 60 percent gross internal rate of return. Twenty-one of its twenty-five buildings became the nucleus of Digital Realty, which by then had a Market Value of about $6 billion.

"That's why we, as contrarians, did very well," Magnuson said. He stepped down as Digital Realty's chairman in 2012, with the oxymoron long since turned into an asset class.

Should we take a fund's word for its own returns? Not entirely. Funds report their own numbers in the best light, and these numbers are GI's own. But the public market set the $6 billion Market Value. Even with a haircut, it was one of the great trades of the decade.

I made a version of the same bet in fiber, as I described in Bandwidth.

Switch—Rob Roy#

"Legend has it that the company managed to acquire what was once meant to be Enron's broadband trading hub for a song," Ashlee Vance wrote in The Register in 2008.

Legend is the right word. The Switch story starts with Enron, and the Enron part has grown in the telling. So let's take the legend apart and see what the sources actually say.

The Enron Building

Bandwidth readers know Enron Broadband Services. My former MFS colleague Scott Yeager became one of its leaders. Enron promised a market where bandwidth would trade like natural gas or electricity. A witness at the 2005 trial of its executives said the software behind that market was called "the secret sauce or pixie dust." The market never arrived. Enron did not survive the attempt.

To trade bandwidth, Enron needed places where many carriers' networks met. One of them was in Las Vegas.

The best period source is a November 2004 report by Ky Plaskon of Nevada Public Radio. According to KNPR, Rob Roy bought a former Enron building about "two years" earlier, at a private Enron sale, for $930,000. Roy was about thirty-three.Switch's October 2017 prospectus gives Roy's age as forty-eight as of September 1, 2017, which makes him about thirty-three in late 2002. KNPR reported his age as twenty-nine in 2004, and The Register gave it as thirty-eight in May 2008. The prospectus is the authoritative source. Enron Broadband Services said it had marketed its properties to 300 IT companies. Its managers said only Roy saw value in this one. Enron's attorneys put it plainly: "It wasn't just an auction. It was a private auction to just Rob Roy."

Roy did not have the money. The closing was delayed while he found financing. A partner called GK Acquisitions came in about two months later. In June 2004, Arden Street Technology Partners bought in for $1.6 million. Then, in October 2004, Tom Thomas of the Las Vegas developer Thomas & Mack bought the building for $3 million and leased it back to Roy. Switch Communications ran the business inside. The building sat on east Sahara Avenue, past U.S. 95.

"Rob Roy is quite lucky because he was prepared when the opportunity came," a Switch spokesman, Mike Ballard, told KNPR.

Data Center Knowledge added detail in 2008. Switch had "acquired a former Enron broadband services facility out of bankruptcy" in December 2002. Enron had planned a commodity bandwidth exchange there and had arranged exceptional connectivity to the site. But the building was small, about 12,000 square feet. And according to court records, the buyer was not Switch or Roy. It was a company called "Colo Gateways."Data Center Knowledge, citing court records. KNPR dates the purchase to "two years" before November 2004; Data Center Knowledge says December 2002. A 2000 Nevada Business article, which I've seen only as a search excerpt, names Roy as chief executive of a Las Vegas colocation start-up called Colocation Gateways.

By 2008, the story had grown. The Register reported that Switch had acquired what was planned as Enron's broadband trading hub, with access to more than twenty carrier backbones in one place. "We were the only ones that bid on it," Roy told Vance. In the same article, Vance flagged a dispute: one of his contacts said the Enron network had gone to a different buyer, Wilhagan Ventures, later WVFiber.

And when Switch finally went public in 2017, its prospectus didn't mention Enron at all. The word appears nowhere in it. The company told its origin story differently: "Mr. Roy first began developing data center facilities in 2000, merging several predecessor companies into Switch, Ltd. after its formation." He had been its CEO and chairman "since 2003." In his letter to investors, Roy put it in one line: "I knew data would run the planet."

So here is what we can say. In late 2002, a thirty-something Las Vegas real-estate man bought a small former Enron Broadband building, wired to many carriers, for about $930,000 at a private bankruptcy sale. He needed partners to close, and he sold the building two years later and leased it back. Was it "Enron's broadband trading hub"? That depends on how big a fish you want the story to be. The 12,000 square feet were real. The carriers were real. The hub grew with each retelling.

The Kid from Wisconsin

Roy grew up in Wisconsin. He helped his paralyzed father earn a degree in computer science. He went from commercial real estate in Las Vegas into colocation. His brother, Mike Borden, became Switch's president.

A real-estate man buying a building cheap makes a real-estate story. What Roy did next made a technology story. He used the carrier connections Enron had paid for as the seed for something much larger.

In 2008, Switch was building the SuperNAP. Data Center Knowledge put it at about 400,000 square feet and $350 million, with room for 7,000 cabinets at full build-out.The Register put the SuperNAP at 407,000 square feet and 1,500 watts per square foot. It was designed to support more than 1,500 watts per square foot with air cooling, backed by a 250-megavolt-ampere substation and 146 megavolt-amperes of generators. Vance wrote that it would use more energy than three mega-casinos. In Las Vegas, that is saying something.

By September 2010, Switch said the SuperNAP had grown to 740,000 square feet, with more than 100 megawatts committed and up to 17.2 kilowatts per cabinet.

Nevada noticed. In 2015, the Las Vegas Review-Journal reported that the state's tax abatements for data centers traced back to a 2004 program designed to win Switch.I have seen this Review-Journal article ("Legislators told data center tax help will benefit Nevada big," 2015) only as a headline and summary; the paper's site blocked automated reading. It concerned SB 170, which offered twenty-year partial abatements for investments of $250 million or more. By the time of the IPO, Las Vegas produced 99.4 percent of Switch's revenue, and Roy and an affiliate would control about 67.7 percent of the vote.

From a $930,000 building to a $350 million one in about six years. And as we'll see in Chapter VII, the Switch story was just getting started.

Managed Hosting Survives—Savvis, Terremark and Rackspace#

Equinix rented the meeting room. Digital Realty rented the building. Both models dodged the hard part of the old Exodus business, which was running the customers' computers for them.

Exodus had died trying. Did that mean the idea was dead?

Savvis

When we last left Exodus's data centers, Cable & Wireless had bought twenty-eight of them out of bankruptcy for $560 million in cash.The figures are from the Exodus (EXDS) bankruptcy disclosure statement of April 2002: Cable & Wireless's subsidiary bought interests in twenty-eight Internet data centers and "paid $560 million in cash upon closing," on February 1, 2002. "No other parties, however, bid on the assets." Press reports at the time, cited in Chapter IV, described the sale as Exodus's best thirty centers for about $570 million. Exodus had operated forty-four centers in June 2001. Nobody else had bid. The new owner did no better than the old one. In December 2003, Cable & Wireless USA, the unit that held the old Exodus business, filed for Chapter 11.

The Gores Group agreed that month to pay $125 million for the assets. Then came an auction under Section 363 of the Bankruptcy Code, with seven bidders. The winner was Savvis, a St. Louis company run by its chairman and CEO, Rob McCormick. In January 2004, the court approved Savvis's bid: $155 million in cash, plus about $12.5 million of assumed liabilities, for most of the assets of Cable & Wireless USA.

The trade press could not resist. "How the mighty have fallen," InternetNews wrote. "The acquirer gets acquired."

Follow the price down. In February 2000, Exodus had a Market Value of more than $20 billion. In February 2002, twenty-eight of its data centers sold for $560 million. In January 2004, fifteen of those data centers, with more than 5,000 customers including Starbucks and Reebok, went for $155 million.

Savvis paid for the deal with help from its private-equity backers, Welsh Carson and Constellation Ventures, and with a classic Phoenix-era trick. It sold its rights to several of the data centers (five, in later press accounts; four plus an office, in Savvis's own 10-K) to Du Pont Fabros Interests for $52 million and leased them back for fifteen years.Light Reading reported the Du Pont Fabros sale-leaseback as part of the plan, along with about $170 million of subordinated notes from Welsh Carson and Constellation Ventures. Savvis announced on March 8, 2004, that it had closed the purchase. Keep that name in mind. It comes back later in this chapter, and it doesn't come back happy.

Savvis expected the combined company to reach about $700 million of annualized revenue by the end of 2004. The Exodus customers, it turned out, had never left the buildings. They had only changed landlords, twice. And as we saw, Savvis had become the largest tenant of the new real-estate company buying those landlords' buildings, Digital Realty.

Terremark

When we last left Manny Medina, he was building a giant network access point in downtown Miami and telling reporters that his "concern regarding the potential use is zero."

The NAP of the Americas opened in interim space on December 30, 2000. The permanent NAP became operational on June 29, 2001. It occupied 120,000 net usable square feet inside a 750,000-square-foot building called the Technology Center of the Americas.Chapter III, from 2000 press reports, described the Miami building as 700,000 square feet; Terremark's 10-K for fiscal 2001 says 750,000. I've used the 10-K. The 10-K also said completion of the NAP was expected in the third quarter of 2001. The consortium members included AT&T, Global Crossing, Level 3, Cable & Wireless and Sprint.

Look at that member list with what we know now. Global Crossing, the anchor tenant, was bankrupt within months. Cable & Wireless was on its way to losing Exodus. Level 3 barely avoided bankruptcy. Terremark's own auditors weren't sure about Terremark. For the fiscal year ended March 31, 2001, it lost $104.0 million, and the auditors wrote that "substantial doubt exists about our ability to continue as a going concern."

Terremark continued as a going concern anyway, and it did so by owning more, not less. In December 2004, it bought the NAP of the Americas building it had been leasing, with a $49 million mortgage and $30 million of notes. In 2007, it bought thirty acres in Culpeper County, Virginia, for what it called the NAP of the Capital Region, and paid $85.0 million for a hosting company called Data Return.These facts come from the capital history in Terremark's Schedule 14D-9 of February 2011. Terremark had gone public in April 2000 by a reverse merger with AmTec, and in June 2009 it sold $420 million of 12 percent secured notes. In May 2009, VMware bought four million Terremark shares at $5.00 each. By 2008, Terremark had become valuable enough that someone wanted to buy it.

The first approach came from an unnamed private-equity fund in October 2007. In April 2008, with the stock at about $6.00, the fund sent a letter implying $9.00 to $10.00 a share. Credit Suisse ran a market check and contacted twenty-one parties. Twelve signed confidentiality agreements, and six bid, between $6.62 and $9.00 a share. The private-equity fund couldn't finance a buyout. A strategic buyer offered $8.75, mostly in its own stock.

On August 8, 2008, five weeks before Lehman failed, the special committee ended the process. The bids were below what it thought Terremark was worth, the financing was in doubt, and the sale was distracting management.Terremark's Schedule 14D-9 dates the decision to August 8, 2008. Data Center Knowledge, reporting in November 2008, tied the end of the process to the September credit freeze. The filing shows the decision came first. The six bids valued Terremark at 6.8 to 8.8 times 2009 EBITDA. Two more private-equity funds bid $10.25 and $10.50 a share in early 2010, and the board turned both down on February 3, 2010. "The (potential) transaction was incredibly disruptive," Medina told Data Center Knowledge. The stock hit $7.50 on September 19 and was at $4.69 by early November, even though quarterly revenue had grown 31 percent, to $59.6 million.

Medina got his sale two and a half years later, and it started on a conference stage. In late summer 2009, he met Kerry Bailey of Verizon Business on a VMworld panel. That December, Verizon's strategy chief, John Diercksen, visited Miami and said Verizon would want "to participate at some equity level." The equity idea faded. The relationship didn't. In May 2010, Verizon became a customer, licensing 25,000 square feet in Miami and Culpeper. In October 2010, Diercksen offered a 35 to 40 percent premium to the $9.90 share price, about $13.37 to $13.86 a share. Medina said his board wanted the "high teens."

On December 13, 2010, Diercksen came back with $19.00. He called it "all-in." Verizon, the 14D-9 says, "would not participate in any competitive bidding or auction process," and it wanted exclusivity. It got it. In January 2011, Verizon agreed to pay $19.00 a share in cash for Terremark, about $1.4 billion of equity value.Verizon's press release described $19.00 as a 35 percent premium. The 14D-9 calls it an "approximately 46% premium" to Terremark's $13.00 share price on December 13, 2010, the day Diercksen made the offer. Verizon also struck Terremark's proposed forty-day go-shop; Terremark got a two-tier break fee instead. Terremark ran thirteen data centers. Verizon already ran more than 220 of its own. Verizon wanted Terremark for what its press release called its "Cloud Strategy." The deal closed on April 11, 2011.

From a $6.62 low bid in 2008 to $19.00 "all-in" in 2010. The credit crisis had cost Terremark a sale and handed it a better one.

Rackspace

Rackspace was the patient one. Its operations began in 1998 as a limited partnership, and it incorporated on March 7, 2000. The same month, it filed to go public. Then the Nasdaq crashed, and Rackspace withdrew the filing.

It took eight years to try again. Rackspace spent them running servers for small and midsize companies and selling the one thing the Gold Rush hosting companies had forgotten to sell: service. The prospectus called it "our commitment to serving our customers, known as Fanatical Support®."

On August 7, 2008, Rackspace priced its IPO at $12.50 a share, near the low end of its $12-to-$16 range. It sold 15 million shares, about $187.5 million worth, and listed on the New York Stock Exchange as "RAX."The offering included 12.7 million shares from the company and 2.3 million from selling holders, and could raise up to $215.6 million with the overallotment. Like Google four years earlier, it used a Dutch auction to set the price. The timing was close to the edge. Lehman Brothers would fail the next month. Rackspace ran eight data centers in the United States and the United Kingdom and had more than 31,000 customers.

Here is the number that separated it from the class of 2000. In 2007, Rackspace booked $362.0 million of revenue and earned $17.8 million. It made money. Exodus, the year before its IPO, had booked $12.4 million of revenue and lost $25.3 million.

The glass house was back again, and this time it turned a profit.

The Telcos Come Calling

Verizon's deal for Terremark set a price for everyone else. Savvis's proxy tells it as a chain reaction. On January 23, 2011, Barclays brought Savvis's CEO, Jim Ousley, together with CenturyLink's Glen Post. Four days later, Verizon announced Terremark. "In the days following this announcement," the proxy says, "Savvis experienced significant appreciation in its stock price." On February 1, Time Warner Cable announced a deal for NaviSite, another hoster, and Savvis's shares rose again.

CenturyLink opened at $38.00 a share. Post told Ousley that "valuation was the principal issue." Savvis refused to open its books for more due diligence until the price went up. On April 12, the two men discussed $40.00, and the data room opened three days later.

In April 2011, three months after Verizon's deal for Terremark, CenturyLink agreed to buy Savvis for $40 a share: $30 in cash and $10 in CenturyLink stock. That was about $2.5 billion of equity and about $3.2 billion including Savvis's net debt.At closing, each Savvis share became $30 in cash plus 0.2479 CenturyLink shares. CenturyLink also prepaid about $546 million of Savvis credit-facility debt. Savvis stayed in St. Louis under CEO Jim Ousley. The deal closed on July 15, 2011. "CenturyLink is now positioned to address complex customer needs with our colocation, hosting and cloud products," CEO Glen F. Post III said.

Bandwidth readers know CenturyLink as the rural telephone company that had swallowed Qwest. Savvis had paid $155 million for Exodus's remains in 2004. Seven years later, CenturyLink paid about $2.5 billion for Savvis.

Look at the pattern. In 2000, the carriers built or bought hosting businesses at the top of the market: Qwest with its CyberCenters, Global Crossing with GlobalCenter, my own Level 3 with its Gateways. Ten years later, the carriers bought hosting businesses again, at prices that looked a lot like another top. In Chapter VII, we'll see the carriers sell their data centers back out, to the investors who had learned to own them.

The 2008 Test#

In March 2008, J.P. Morgan Chase rescued Bear Stearns. In September, Lehman Brothers went bankrupt. As I wrote in Bandwidth, the Great Recession hit just as my new company, Zayo, was being launched. When we pitched our lenders at CIT in New York, few questions were asked, and the mood was somber. Was our timing bad, or was it an opportunity?

The data-center business faced the same question. In 2001, its collapse had come from the demand side. The forecasts had been wrong, and the buildings were empty. In 2008, the collapse came from the money side. The buildings were full, and the banks were broken. Which would matter more?

The Builder

DuPont Fabros Technology was the first to find out. Readers will remember the Du Pont Fabros name from the Savvis deal, when Du Pont Fabros Interests bought several of the old Exodus centers and leased them back. DuPont Fabros Technology had gone public in October 2007 at $21 a share, as a developer of big wholesale data centers in northern Virginia and elsewhere. Its business was building new space and leasing it to large tenants. Building new space takes borrowed money.

In November 2008, DuPont Fabros told investors it had planned to borrow $300 million to $400 million against one of its Virginia buildings, a project called ACC4 that had been appraised at $680 million. It got $100 million. The company mothballed a $270 million project in Santa Clara and suspended its dividend. Its chief financial officer, Mark Wetzel, explained why the money had dried up: "several banks we were in discussion with have merged or disappeared."

The shares fell 45 percent in a day, to $3. Eleven days later, after talks on a $150 million mezzanine loan ended when the lender changed its terms, DuPont Fabros halted construction in New Jersey and Virginia as well. "After a careful evaluation of our development initiatives we believe this is the prudent course of action at this time," CEO Hossein Fateh said. The shares fell to $1.93 in early trading.Santa Clara had been halted on October 28, 2008. The November 17 halt covered NJ1 in Piscataway, New Jersey, and ACC5 in Virginia. DuPont Fabros said it was negotiating a $150 million loan from a pension fund when it reported in early November; that is the mezzanine financing whose terms the lender later revised. At $3 a share, the company's Market Value was $101.5 million.

From $21 to $1.93 in thirteen months. That is a Bust-sized drop. But look at what the reports didn't say. None of them said the tenants were leaving. DuPont Fabros had run out of lenders. It still had its customers.

The Landlord

Digital Realty had tenants on long leases and buildings it had bought cheap. In March 2009, Data Center Knowledge quoted the real-estate trade paper GlobeSt: "Digital Realty Trust is living in a different world than most of the rest of the commercial real estate industry."

The numbers backed it up. The portfolio had grown to 11 million square feet and was 95 percent leased. New leases were being signed at three times the rents of the leases expiring. In the fourth quarter of 2008, as Lehman's collapse rippled through the economy, Digital Realty leased 276,000 square feet, to tenants that included IBM, Cisco, Cigna and NTT America.

"We haven't seen any material change at all to our experience with our prospective customers," Foust said.

Digital Realty was the best-performing data-center stock of 2008. In that year, holding still counted as winning.

The Operator

Equinix grew right through it. Its 2008 revenue was $704.7 million, up 68 percent. It earned $131.5 million, helped by a one-time tax benefit.The 2008 net income included a $104.9 million tax benefit in the fourth quarter. Fourth-quarter revenue was $190.7 million, up 4 percent from the third quarter. It ended the year with $307.9 million of cash and investments and guided to $855 million to $875 million of revenue for 2009.

"Equinix delivered exceptional results in 2008, creating a strong platform for continued growth in 2009," Smith said.

Five years earlier, Equinix had booked $117.9 million of revenue and lost $84.2 million. Now its revenue was six times larger, and the company was profitable in the year Lehman failed.

Even Terremark's lost summer proved the point. Its sale died partly because a buyer couldn't finance it. Its revenue kept growing.

The Verdict

In 2009, the market caught up. Data Center Knowledge tracked the sector's stocks and found that every one of them gained at least 53 percent that year. "Share prices of many key players in the sector finished with triple-digit gains for 2009," it wrote. DuPont Fabros, the one that had nearly run out of money, led the group. NaviSite, Rackspace, Switch and Data, Savvis and Equinix all more than doubled. Digital Realty rose about 53 percent, the smallest gain, because it had never fallen very far.

What should we take from the 2008 test? The Bust of 2001 had taught that buildings survive and balance sheets don't. The crisis of 2008 added a second lesson. A building full of real tenants on real leases survives even when the banks don't. The companies that suffered were the ones still building with borrowed money. The ones that owned finished buildings, full of customers who couldn't easily leave, barely noticed.

Cash flow was king. It always is.

The Chorus Returns

The analysts noticed. In 2001 and 2002, Tier 1 Research had counted the empty floors. In June 2010, the same firm's Antonio Piraino told Data Center Knowledge, "Demand is far outweighing supply." The brokers at Grubb & Ellis could find only six single-story buildings in the country with more than 30,000 square feet of raised floor built out and available.

The skeptics had history on their side. Data Center Knowledge reminded its readers of the old strategy of "build it, and they will come," and of the years from 2001 to 2003, when WorldCom and AboveNet built "barns" of more than 100,000 square feet at about $100 million apiece. Was this the same movie? Piraino said no. "We do not see oversupply in this market," he told a hosting conference at the Bellagio in Las Vegas in September 2011. The difference, he said, was the balance sheets: "Companies were incredibly leveraged. Now people are tightening their belts." The new buildings went up in pods of 8,000 to 10,000 square feet, a phase at a time, as tenants signed.

This time the bull was right. The chorus doesn't always sing the wrong part.

Notice what was missing from the Phoenix, too. Nobody in this chapter raised money on a forecast. Equinix bought finished buildings for a fraction of what they had cost. GI did the same and leased its buildings to companies like IBM and Cisco. Rackspace waited until it was profitable to go public. In the Gold Rush, as we saw in Chapter III, the forecast was the asset. In the Phoenix, the lease was the asset. When we get to the AI forecasts at the end of this book, that difference will be worth remembering.

From Phoenix to Cloud#

Let's run the Bandwidth scorecard on the Phoenix years.

Equinix went from 30 cents a share and a Singapore rescue in 2002 to eighty-seven centers in thirty-five metros in 2010. GI Partners turned $280 million of equity and twenty-five unwanted technology buildings into Digital Realty, a company the market valued at about $6 billion. Rob Roy turned a $930,000 Enron leftover into a $350 million SuperNAP. Savvis bought fifteen of Exodus's data centers for $155 million and sold the company to CenturyLink for about $2.5 billion of equity. Terremark went from a going-concern warning to a $1.4 billion sale to Verizon. Rackspace went public eight years late, and profitable.

The Phoenix had risen. The buildings that the Gold Rush paid for were full, and the people who bought them in the wreckage had made fortunes.

But look again at the customer lists. IBM, Cisco, Cigna, Starbucks, Reebok, 31,000 small businesses at Rackspace. These were companies renting space and servers one rack at a time. The biggest buyers of computing in the world hadn't shown up yet as tenants. They were about to show up as something else.

On August 24, 2006, while Equinix was waiting for Steve Smith and Digital Realty was buying AboveNet's buildings, Amazon opened the beta of a service it called the Elastic Compute Cloud. BusinessWeek reported that it cost about 10 cents per server-hour. "Amazon EC2 changes the economics of computing by allowing you to pay only for capacity that you actually use," Amazon said. Tymshare had sold a version of the same idea in the 1960s.

The glass house was coming back one more time. This time it would rent itself out by the hour, from buildings bigger than anything Exodus ever imagined. And it would change who owned the data center, who paid for it and who called the shots. The table was set for the cloud.

At 2 a.m. on August 24, 2006, Amazon opened a test of something it called the Elastic Compute Cloud. Within about five hours, every test slot had been claimed.The test was a limited beta. Amazon's own announcement is dated August 24, 2006. Some secondary timelines give August 25, probably the next day's news cycle. A Slashdot item the same day said the beta appeared to be limited to existing S3 users and was "already full."

The people who grabbed those slots weren't renting a cage in Santa Clara or a rack in Ashburn. They weren't signing a three-year contract, buying servers or waiting on a utility to deliver power. They were renting a computer from an online bookstore, by the hour, on a credit card. BusinessWeek put the price at about 10 cents per server-hour.

Wall Street wasn't impressed. Amazon was spending heavily on technology, and analysts couldn't see when the spending would pay off. In November, BusinessWeek ran its cover story under the headline "Jeff Bezos' Risky Bet."

The bet would change the data-center business more than anything since the Bust. Over the next ten years, the cloud created the largest customers the industry had ever seen: Amazon, Google, Microsoft and Facebook. Those customers then discovered they could build their own buildings, and they started doing it in places nobody in the hosting business had ever heard of: a river town in Oregon, a potato town in Washington, a county seat in Oregon whose whole electric load was smaller than one building. What they didn't build, they leased by the tens of megawatts, much of it in one county in Northern Virginia. A new breed of developer grew up to serve them, built huge buildings on spec and took them public as real-estate investment trusts. And a fiber company in Boulder, Colorado, decided it should own some buildings too.

When we last left the data-center business, the Phoenix had risen. Equinix, Digital Realty and the survivors had bought the Bust's empty floors at pennies on the dollar and filled them with tenants. Now the tenants were about to change, and so was the balance of power between the people who own the buildings and the people who fill them.

Amazon Web Services—Andy Jassy#

Andy Jassy's job at Amazon in the early 2000s was to follow Jeff Bezos around. He was the chief of staff, and from that seat he noticed something odd. Every time an Amazon team started a new project, it rebuilt the same infrastructure from scratch.

In 2003, at an executive retreat at Bezos's house, the leadership team sat down for what was supposed to be a thirty-minute exercise on Amazon's core competencies. It ran long. By that summer, the idea had become a plan to build an "operating system" for the Internet that outsiders could rent. "In retrospect it seems fairly obvious," Jassy said in 2016, "but at the time I don't think we had ever really internalized that."Jassy's account comes from a 2016 retrospective in TechCrunch, not a period source. The BusinessWeek cover story of 2006 doesn't mention him.

The first product was storage. On March 14, 2006, Amazon announced the Simple Storage Service, S3, with Jassy identified as vice president of Amazon Web Services. The pricing fit on a napkin: 15 cents per gigabyte per month to store data and 20 cents per gigabyte to move it. No minimum fee. Objects could be as small as one byte or as large as five gigabytes, and you could store as many as you liked. "Amazon S3 enables developers to focus on innovating with data, rather than figuring out how to store it," Jassy said in the press release.

Five months later came EC2, the rentable computer. Amazon's announcement was a short post with no prices in it, but one sentence captured the whole business model: "Amazon EC2 changes the economics of computing by allowing you to pay only for capacity that you actually use."

I have read that sentence many times, and every time I hear an echo. In Chapter I, Tymshare and its competitors rented time on a mainframe to customers who couldn't justify buying one. You dialed in, you computed, you paid for the minutes. Then the PC arrived, and everyone decided they'd never rent a computer again. Forty years later, an online retailer was renting computers by the hour to customers who didn't want to own one.

The glass house was back again. This time you could rent it with a credit card.

The Risky Bet

BusinessWeek's cover story that November laid out both sides of the argument. On the bull side were the early customers. SmugMug, a photo-sharing site, said S3 was saving it about $500,000 a year. Powerset, a search startup, expected to cut its first-year capital costs by more than half against a hardware budget of more than $1 million. SmugMug's Chris MacAskill had become a true believer: "Everything we can get Amazon to do, we will get Amazon to do."

On the skeptic side was Amazon's income statement. Spending on technology and content had risen 52 percent to $485 million, and analysts saw no near-term return. Why was a retailer with thin margins pouring money into servers for other people's startups?

Bezos didn't promise much. "We think it's going to be a very meaningful business for us one day," he told the magazine.

He was underselling it. By 2016, ten years after the S3 launch, AWS was running at about $10 billion of revenue a year. Jassy, looking back, sounded almost as surprised as the analysts: "I don't think any of us had the audacity to predict it would grow as big or as fast as it has."

Why It Mattered to the Buildings

For the data-center business, AWS changed who the customer was. In the hosting Gold Rush, the customer was a dot-com that rented a cage from Exodus and filled it with its own servers. In the Phoenix years, the customer was an enterprise or a carrier that leased space and power from Equinix or Digital Realty. Every one of those customers was small, and there were thousands of them.

AWS turned many of those small customers into one big one. SmugMug didn't need a cage anymore. Amazon needed the buildings instead, and Amazon would need them by the dozen. The same was about to be true of Google, Microsoft and Facebook. The customers who used to rent space were starting to rent computers, and the companies renting out the computers were becoming some of the biggest buyers of data-center space anywhere.

In Bandwidth, I speculated that one of the Big Tech companies might someday decide that owning a national fiber network was a strategic addition to its cloud, "not unlike Amazon's AWS play." That's how big a shadow AWS casts over infrastructure. A business that started as a way to stop Amazon's own engineers from reinventing the wheel became the template for how every piece of digital infrastructure gets valued.

Google, Microsoft, Yahoo and Facebook—Building Their Own#

Two months before EC2 opened its doors, John Markoff and Saul Hansell of the New York Times went to The Dalles, Oregon, a town on the Columbia River about 80 miles east of Portland. In June 2006, they reported on what Google was building there under the code name "Project 02."

The complex was about the size of two football fields, with twin cooling plants four stories tall. Google held a permit for a third building. The town's city attorney and city manager had signed confidentiality agreements with the company, so they couldn't say much. The newspaper ran the story under the headline "Hiding in Plain Sight, Google Seeks More Power."

Why The Dalles? Two reasons, and both should sound familiar to anyone who read Bandwidth. The first was cheap hydropower from the Columbia River dams. The second was surplus fiber left over from the dot-com bust. The overbuilt networks of the Boom were finally finding a use, and the use was a search engine's computer room in a town most Americans couldn't find on a map.

The land was cheap too. In February 2005, a company called Design LLC, buying for Google, paid $1.87 million for 30.19 acres in the Port of The Dalles's industrial park, with an option on 80 more.The 2005 purchase is known to me only from an excerpt of a GlobeSt report ("Google buying 30 acres on Columbia River," February 18, 2005), and the 2016 Oregonian story only from an excerpt of its reprint in the Spokesman-Review. I haven't found The Oregonian's 2005 coverage. The port expected 50 to 100 jobs. Eleven years later, The Oregonian reported that Google had bought the site of the old Northwest Aluminum smelter and that the port's industrial land was nearly gone. On the Columbia, the servers had replaced the smelters.

Danny Hillis, the computer scientist, gave the Times its best line: "Google has constructed the biggest computer in the world, and it's a hidden asset."The same New York Times article offered a "best guess" that Google had "more than 450,000 servers spread in at least 25 locations around the world." Google's tax deal with Wasco County and the city of The Dalles, two fifteen-year agreements, the first signed in 2005, is known only from search snippets of later coverage. When the exemptions ran out about fifteen years later, Google reportedly received a $5.2 million tax bill on a facility the county valued at $335 million.

I had my own reason to notice. In 2006, I was recovering from double hip-replacement surgery after leaving ICG, and I passed the long days by calling the CEOs of fiber companies. The premise for Zayo came out of those calls. As I wrote in Bandwidth, the first plank was that the Insatiable Demand for Bandwidth would hold for generations, and the most visible proof was "the daily widespread use of the Google search engine and Google Maps, the monumental adoption of streaming, and the continued expansion of cloud services and data centers." The second was that the fiber business had Too Many Fiber Providers, not too much fiber. Google's building on the Columbia was the first plank in concrete. Somebody had to connect it.

The Times also noticed who was chasing Google. About 130 miles north, Microsoft and Yahoo were building in Wenatchee and Quincy, Washington. "But it is a race in which they are playing catch-up," the reporters wrote.

Potato Country

Three months earlier, in March 2006, the Associated Press had visited Quincy and found a farm town of about 5,300 people that was about to get very rich. Microsoft had a tentative deal for 74 acres in an industrial park at about $1 million. Yahoo had a tentative deal for 50 acres at about $500,000, plus a roughly $6 million, ten-year lease in nearby Wenatchee. The two companies could roughly double Quincy's tax base of about $800 million.

The locals were thrilled. "This could be a real boon to Quincy and to Grant County," said Curt Morris of the Port of Quincy. The city administrator, Tim Snead, was even happier: "It's been a quiet little town for many, many years, but it's going to be very exciting."

It got exciting. Six years later, in September 2012, James Glanz of the New York Times went back to Quincy for a series called "The Cloud Factories." The headline was "Data Barns in a Farm Town, Gobbling Power and Flexing Muscle."

The numbers told the story. Microsoft had bought about 75 acres of former bean fields and asked for a 48-megawatt substation. Together, Microsoft and Yahoo drew 41.8 megawatts. All the residential and small commercial accounts combined averaged 9.5 megawatts. The companies were paying 2.5 to 3.8 cents a kilowatt-hour, against a national industrial average of 6 to 7 cents.

Then came the part nobody in Quincy had imagined in 2006. Under its power contract, Microsoft faced a $210,000 penalty because its actual use didn't match its estimate.The source summary I worked from describes the penalty as being for "underestimating" Microsoft's use, but the rest of the account, in which Microsoft raised its consumption to shrink the penalty and called that "unnecessarily wasteful," implies that actual use fell short of the estimate. Glanz's own text has it both ways: the penalty was "for underestimating its power use," and Microsoft was being penalized "for using too little power." To shrink the penalty, Microsoft raised its consumption by 5 to 7 megawatts, burning power it didn't need. Microsoft itself called the episode "unnecessarily wasteful." The utility board waived all but $60,000 of the penalty. Yahoo paid its own penalty, $94,608, in full. Glanz also found that Microsoft's backup diesel generators had run 3,615 hours in 2010. Yahoo's ran 65.

Microsoft's spokeswoman called it "a one-time event that was quickly resolved." Fair enough. But the episode captured something that would matter a great deal by the end of this book. The new cloud customers were so big that a single company's rounding error was several times the load of a whole town. And the town's leverage over the company was about $150,000 of waived penalties.

Washington Notices

Quincy wasn't the only place counting megawatts. On December 20, 2006, four months after EC2 opened, Public Law 109-431 was signed. The short statute directed the Environmental Protection Agency to report on "the rapid growth and energy consumption of computer data centers." For the first time, Congress had formally noticed the buildings.

EPA delivered its Report to Congress on August 2, 2007. In 2006, it found, the country's data centers had used about 61 billion kilowatt-hours, 1.5 percent of all U.S. electricity, at a cost of about $4.5 billion. That was "more than the electricity consumed by the nation's color televisions," and about as much as 5.8 million households. Peak demand was about 7 gigawatts, "equivalent to the output of about 15 baseload power plants." Use had more than doubled since 2000, and about half of it went to power and cooling equipment, not to the computers.

Then EPA made a forecast. Under current trends, it said, data-center use "could nearly double again" by 2011, to more than 100 billion kilowatt-hours a year and $7.4 billion, with peak demand rising to 12 gigawatts. That was about ten more power plants. Best practices, the agency added, could cut that 2011 figure by up to 45 percent.

Hold on to that forecast. We'll check it at the end of the chapter, and again at the end of the book.

Opening the Blueprints

Facebook came late to building, and when it did, it gave the blueprints away.

Facebook started out as a tenant. As we'll see shortly, one of its landlords was DuPont Fabros in Ashburn. Then it built its own, in Prineville, Oregon, in Crook County. Construction started in early 2010, and the building began operating in April 2011. Phase one cost $210 million, covered 300,000 square feet and drew 28 megawatts. All the other homes and businesses in Crook County combined used about 30 megawatts. The campus could eventually reach about 78 megawatts. The site employed fifty-five people at an average salary of $47,200.Square-footage figures for Prineville vary by source: 147,000 (planned), 150,000, 300,000 and 330,000. The 300,000 figure is from Data Center Knowledge's January 2012 report on the $210 million phase one. The exact opening day in April 2011 is unconfirmed.

On April 7, 2011, Facebook did something none of the other giants had done. It published the designs of the Prineville building and the servers inside it, including the power supplies and the battery backup, and called the effort the Open Compute Project. AMD, Dell, HP and Intel had helped develop the technology. Dell and Synnex would build and sell hardware to the specification.

The efficiency numbers explained why Facebook bothered. Prineville used 38 percent less energy to do the same work as Facebook's earlier facilities and cost 24 percent less to build. Its power usage effectiveness, the industry's measure of how much electricity goes to overhead rather than computing, was 1.07, against 1.5 in its existing facilities.Power usage effectiveness (PUE) is the total power a data center draws divided by the power its computers use. A PUE of 1.5 means half a watt of overhead, mostly cooling, for every watt of computing. A PUE of 1.07 means seven-hundredths of a watt. For comparison, the Uptime Institute's surveys put the industry average at about 2.5 in 2007, 1.98 in 2011, 1.58 in 2018 and 1.54 in 2025. I have seen those figures only in excerpts of Uptime's reports.

Jonathan Heiliger, Facebook's vice president of technical operations, made the pitch: "We think it's time to demystify the biggest capital expense of an online business — the infrastructure." Mark Zuckerberg added, "We're not the only ones who need the kind of hardware we're building out." And Amir Michael, one of the hardware engineers, explained why giving it away cost Facebook nothing: "We don't hold this technology as a key to our business."

Think about that last line. To the hosting companies of 1999, the building was the business. The palm scanners, the cooling, the generators: that was what Exodus sold. To Facebook in 2011, the building was a cost center to be engineered down and then shared, so that suppliers would make the parts cheaper for everyone. When your biggest potential customer tells you your product isn't a competitive advantage, you should pay attention.

The Cloud Buys the Carrier Hotel

The giants didn't only build in farm towns. In December 2010, the Wall Street Journal reported that Google had signed a contract to buy 111 Eighth Avenue in Manhattan for about $1.9 billion.The price was never officially disclosed. Data Center Knowledge, citing the Wall Street Journal, reported about $1.9 billion. TechCrunch and Fox Business reported about $1.8 billion, and the Motley Fool said the price could reach $1.9 billion with transfer taxes and fees. The contract was signed around December 2–3, 2010, and the deal closed on December 22, 2010. The sellers were Taconic, the New York State Common Retirement Fund and Jamestown.

111 Eighth was, in Data Center Knowledge's words, "one of the premier 'carrier hotels' in New York." Built in 1932 as the Port Authority Commerce Building, it fills a full city block, about 2.9 million square feet, from 15th Street to 16th Street between Eighth and Ninth Avenues. Taconic converted it to telecom use in the late 1990s. By 2010, data-center and telecom uses took up roughly a quarter of the building. Digital Realty, Equinix and Telx were colocation tenants. Google already occupied about 500,000 square feet as offices.

So one of the companies that had created the cloud now owned one of the buildings where the Internet's networks met. The landlord of Equinix and Digital Realty in Manhattan was a search engine.

Build or Lease?

By 2011, the largest cloud companies had answered the question that hung over the whole industry: would they build or lease? The answer was both. They built where land and power were cheap and they could take their time, in The Dalles, Quincy and Prineville. They leased where they needed to be close to networks and customers, or where they needed space faster than they could build it.

That answer cut both ways for the landlords. The bad news was that the biggest customers knew exactly what a building cost to build, because they built them. Facebook had published its blueprints to prove it. The good news was that they couldn't build fast enough to keep up with their own growth, and for that they needed somebody else's building. Nowhere was that need bigger than in one county outside Washington, D.C.

Data Center Alley—Loudoun County#

As I wrote in Bandwidth, WorldCom emerged from bankruptcy in April 2004, changed its name back to MCI and moved its headquarters to Ashburn, Virginia. Its $41 billion of debt had been cut to $5.5 billion. A year later, Verizon bought it for $6.6 billion.

The move was a fitting end for the company that had carried so much of the Internet's early traffic. Northern Virginia had been an Internet town from the start. Rick Adams ran UUNET from his house there, and his conversations with MFS's Scott Yeager sparked the creation of MAE-East, one of the Internet's first great meeting points. By the time MCI moved in, Ashburn was already the place where networks went to meet. Equinix had built its first Ashburn data center, DC1, in 1998, and by 2011 it had added four more, totaling nearly 500,000 square feet.The Equinix square footage is from the 2011 Washington Post story as summarized by Data Center Knowledge. As we saw in Chapter IV, AboveNet sold its own Northern Virginia data centers out of bankruptcy for $63 million.

What happened next turned a suburb into the center of the cloud.

In August 2011, the Washington Post reported that Loudoun County had about forty data centers with roughly 4 million square feet of space. The newspaper compared it to twenty-two Wal-Mart Supercenters. Another 800,000 square feet were planned, and the county expected 6.5 million square feet by 2021.I haven't read the Washington Post original. Its content comes from Data Center Knowledge's August 28, 2011 summary, The headline is "After dramatic growth, Ashburn expects even more data centers." Its web address carries the date June 9, 2011, but the page's metadata gives August 27, 2011, and Data Center Knowledge described it as "a story this weekend."

The county's forecast was way too low. In September 2012, Buddy Rizer, then Loudoun's assistant director of economic development, counted nearly 5 million square feet, with 2 million more in the pipeline. He projected 10 million square feet within five years. Digital Realty saw more than 40 megawatts of potential demand in Northern Virginia against about 27 megawatts of supply. Philip Sandino of Dominion Virginia Power said his utility was "seeing an amazing number of inquiries (from data center developers) regarding power."

Rizer's projection was too low too. By September 2016, Loudoun had more than sixty facilities covering about 10 million square feet, with another 3.5 million under construction. The county had hit its 2021 forecast five years early, and then some. According to Rizer, the county hadn't gone a single day without data-center construction since 2008.

Why Loudoun? The power was cheap, about 20 percent below average. The networks were already there. And the county made it easy, because the tax money was extraordinary. By 2016, data centers were paying Loudoun about $110 million a year in taxes. "The tax benefits to Loudoun County have been incredible," Rizer said. Supervisor Ron Meyer put it more simply: "There's 21st-century technology in those buildings."

That same September, the Washingtonian ran a story under the headline "70 Percent of the World's Web Traffic Flows Through Loudoun County." It's a great headline. I can't find a source for it. The magazine presented the figure as an estimate without citing anyone, and later outlets traced it to the county's own economic-development office.Later outlets, including Virginia Business in 2017, attribute the 70 percent figure to Loudoun's Department of Economic Development. I've seen those only as search snippets. Treat it the way you'd treat a forecast in a 1999 IPO prospectus.

The Forecast Ran the Other Way

Here is what strikes me about Loudoun. In Chapter IV, we watched the hosting forecasts come down: eighty-fold growth in five years, then $17 billion by 2003, then about $9 billion. The buildings went up on the forecasts, and the bankruptcies came when the forecasts came down.

In Loudoun, the forecasts kept coming up short. The county's 2011 forecast for 2021 was beaten in 2016. Rizer's 2012 projection was beaten too. For the first time in this book, demand for data-center space was outrunning the people who were paid to predict it.

The brokers' numbers tell the same story. According to JLL, Northern Virginia leased 62 megawatts of data-center space in 2015. In 2017, it leased 115 megawatts, a record. No other market in the world had ever leased more than 59 megawatts in a single year. Land in Ashburn was going for more than $1.2 million an acre. "In Northern Virginia, supply creates demand, and demand drives supply," JLL's Allen Tucker told a forecast conference in Ashburn in early 2018.Reported by Rich Miller in Data Center Frontier, February 2, 2018, from JLL's presentation at the CAPRE Data Center Forecast in Ashburn. The article also gives a 2016 figure of 113 megawatts, though its text has typos. JLL credited Digital Realty with 62 percent of Northern Virginia's 2017 leasing and CyrusOne with 21 percent.

That sentence would have made a fine slogan for Exodus in 1999. Supply creating its own demand is how a bubble talks about itself. What made Ashburn different was who signed the leases.

That's what a handful of very large customers does to a market. When Exodus was selling to thousands of dot-coms, the demand was spread thin and evaporated fast. When Amazon, Microsoft, Google and Facebook needed space, each one needed it by the tens of megawatts. Each lease was enormous. And the developers who could deliver those leases fastest were about to become the hottest stocks in real estate.

In 1998, Andrew Odlyzko showed that the forecasts for Internet traffic were far too high. Nobody had to debunk Loudoun's forecasts. The buildings did it for them, by filling up faster than the county predicted. Keep both lessons in mind for the end of this book. Forecasts can be wrong in either direction, and the direction depends on who the buyers are and whether they can pay.

The Wholesale Developers—Hossein Fateh, Chad Williams and the New REITs#

The data-center business had always sold space by the cage, the rack and the cross-connect. That was Equinix's game, and it's called retail colocation. The cloud customers didn't want a cage. They wanted a whole building, or a whole floor of one, delivered with power and cooling, on a long lease. That's wholesale.In wholesale, a tenant typically leases a large block of space with its own dedicated power, often a megawatt or more, on a lease of several years, and brings its own equipment and staff. In retail colocation, a tenant rents a cage or a few racks, buys power by the circuit and pays for cross-connects to other tenants. Equinix is the archetype of retail; DuPont Fabros was the archetype of wholesale.

The wholesale developers built big, powered buildings and leased them in big pieces to a small number of very large tenants. It was closer to office real estate than to hosting, and Wall Street learned to value it the same way: as a real-estate investment trust. Digital Realty had invented the data-center REIT with its 2004 IPO, as we saw in Chapter V. Over the next decade, four more followed it to market.

DuPont Fabros

DuPont Fabros Technology was the first. In October 2007, it priced 30.5 million shares at $21, the top of its range, and raised about $641 million. Lehman and UBS led the deal. The shares traded as high as $23 early. DuPont Fabros became the second data-center REIT after Digital Realty.The prospectus cover gives 30,500,000 shares; Data Center Knowledge reported 30.575 million. At $21 a share, 30.5 million shares come to about $640.5 million. DuPont Fabros's 10-K for 2007, seen only as a search snippet, says the IPO closed on October 24, 2007, with 35,075,000 shares including the underwriters' over-allotment, for $736.6 million gross and $677.0 million net.

Read the prospectus, and you can see the whole decade coming. Two tenants, Microsoft and Yahoo, paid 85.8 percent of DuPont Fabros's rent. UBS, Google and Fox Interactive Media trailed far behind. The company's new prototype in Ashburn, ACC4, had a first phase in service and fully leased, and its second phase was 56.3 percent pre-leased before it opened. And the risk factors said out loud what every landlord in this chapter would come to fear: "our tenants, including Google, Microsoft, and Yahoo!, own or may in the future develop their own data centers."

The founders did fine. For their property interests and management contracts, Hossein Fateh received units and cash worth about $253 million at the IPO price, and his partner, du Pont, about $275 million.Per the prospectus, Fateh received 6.53 million operating-partnership units plus $37.0 million in cash, together valued at $174.1 million, for his property interests, and units worth $79.2 million for management contracts. Du Pont received 7.58 million units plus $37.0 million in cash, valued at $196.2 million, plus $79.2 million in units. Together they would hold about 35.4 percent of the company. The prospectus states ACC4's first phase came into service in July 2007, 100 percent leased; the 56.3 percent pre-leasing figure for phase two is as of October 15, 2007. The prospectus does not name Facebook.

Data Center Knowledge added a tenant that the prospectus doesn't name. Part of the Ashburn space, it reported, was leased "after Facebook signed a lease for data center space." The social network that would later publish its own blueprints was, in 2007, a wholesale tenant in Loudoun County.

As we saw in Chapter V, DuPont Fabros then walked straight into the financial crisis. A year after the IPO, its shares were trading below $2. "Several banks we were in discussion with have merged or disappeared," its CFO said. CEO Hossein Fateh halted construction in New Jersey and Virginia and mothballed a $270 million project in Santa Clara.

And then the cloud bailed it out. In 2009, every data-center stock tracked by Data Center Knowledge gained at least 53 percent, and DuPont Fabros led the group. "Share prices of many key players in the sector finished with triple-digit gains for 2009," the trade publication reported. The tenants had kept paying rent through the credit crisis, and investors noticed.

Let's do the arithmetic. DuPont Fabros went public at $21 in October 2007. Thirteen months later, its shares traded at $1.93, a drop of more than 90 percent. That was an Exodus-sized collapse. The difference was what sat underneath it. Exodus's buildings were full of dot-coms that couldn't pay. DuPont Fabros's buildings were full of tenants like Facebook that could. The problem was the financing, not the demand, and when the credit markets came back, so did the stock. The buildings were fine. The buildings were always fine. This time, so were the tenants.

CoreSite and Vantage

In 2010, two more platforms appeared, both backed by private equity.

CoreSite was the odd one in the group. Its data centers, its prospectus said, "have been historically funded and held through real estate funds affiliated with The Carlyle Group," and "the first data center in our portfolio was purchased in 2000," at the top of the Gold Rush. In 2007, Carlyle's funds went shopping: 427 South LaSalle in Chicago for $35.0 million, a business park in Santa Clara for $37.8 million, 70 Innerbelt in Boston for $32.5 million, and a building in Reston, Virginia, for $45.0 million. CoreSite had more than 600 customers, more than 200 of them network providers. That made it more like Equinix than like DuPont Fabros, a place where networks met.

CoreSite went public in September 2010, selling 16.9 million shares at $16, the midpoint of its range. It closed its first day at $16.01 and raised about $270 million. It had eleven data centers with about 2 million rentable square feet, and it was based in Denver, just down the road from Zayo's headquarters in Boulder.

Vantage Data Centers launched three months earlier, in June 2010, with funding from Silver Lake. Its first campus covered 18 acres on Walsh Avenue in Santa Clara, served by dual 50-megawatt substations. It already operated more than 6 megawatts of space there, with 6 more under construction. Silver Lake didn't disclose how much it put in. Remember the name Vantage. It will change hands in the next chapter.

CyrusOne and QTS

The last two arrived in 2013, and both were stories of a different kind of owner deciding to get out of the way.

CyrusOne came from Cincinnati Bell, the old phone company, which had run a data-center business in Cincinnati for more than ten years. In June 2010, it bought a Texas operator called Cyrus Networks for about $526 million, net of cash acquired.The CyrusOne prospectus gives "approximately $526 million, net of cash acquired, in June 2010." Cincinnati Bell's May 13, 2010 Form 8-K, announcing the deal, gave a $525 million cash price before adjustments for debt and fees. Cincinnati Bell had also bought GramTel, with centers in South Bend and Chicago, for about $20 million in December 2007. Cincinnati Bell's IPO release, which I read, gives the share count, price and ownership. The total raised, $313.5 million, also comes from snippets. In January 2013, Cincinnati Bell took CyrusOne public at $19 a share, selling 16.5 million shares on Nasdaq under the ticker CONE. Cincinnati Bell kept about 72 percent. A phone company had figured out that the market would pay more for its data centers as a separate REIT than as a line item inside a telco.

CyrusOne's tenants weren't the cloud companies. The energy industry paid 38 percent of its rent, "including five of the six 'supermajor' oil and gas companies," which needed enormous computing power to process seismic data. Nine of the Fortune 20 were customers. The cloud was eating the building, but in Texas the oil companies were still renting space the old way.

QTS Realty Trust priced its IPO in October 2013 at $21 a share, raising $257.25 million. That was well below the $27 to $30 range it had marketed, according to Renaissance Capital.The marketed range is from Renaissance Capital. QTS's own release, which I read, gives the price and the gross proceeds. Its chairman and CEO was Chad Williams, and his prospectus read like a Phoenix-era treasure hunt.

Williams had spent twenty-four years developing what the filing called "special-use, mission-critical facilities." In 2003, he "began acquiring data centers," starting from "a single facility." His method, in the prospectus's words, was "obtaining and revitalizing large, scalable properties." In 2006, he bought a former distribution center for a major retailer near Atlanta, with a 72-megawatt Georgia Power substation on the site. By 2013, it was QTS's largest data center. In 2010, he bought a building in Richmond, Virginia, "through a bankruptcy process." Its former owner, a semiconductor maker, had invested, by QTS's estimate, "over $1 billion" in it, and the building came with 110 megawatts of utility power. In February 2013, QTS bought another old chip plant, in Dallas: 698,000 square feet on thirty-nine acres, with 140 megawatts of power feeds. The cash it paid was $21.2 million.

A chip maker's billion-dollar building, bought out of bankruptcy, and a second fab with 140 megawatts for $21 million. Fiber wasn't the only thing that sold for pennies on the dollar. Power did too, if it came wired into a failed factory, and Williams went shopping for it. The market wasn't sure yet what a data-center REIT was worth. It would figure that out in the next chapter.

A Landlord's Dilemma

By 2013, five data-center REITs were trading in the United States: Digital Realty, DuPont Fabros, CoreSite, CyrusOne and QTS. Equinix would convert to a REIT too, operating as one from January 1, 2015.

The wholesale model had one big strength and one big weakness, and they were the same thing. The tenants were the best credits in the world. Facebook, Microsoft and the rest were not going bankrupt, unlike the dot-coms that had filled Exodus's cages. But there were only a handful of them, and each one knew exactly what its leased space would cost to build, because each one was building its own. A landlord with five customers who can all build without you doesn't have much pricing power. A landlord with five customers who can't build fast enough has a great business.

Which of those was the wholesale business? For most of the decade, the answer was the second. But nobody could be sure the answer wouldn't change.

Zayo—zColo#

In December 2010, Google became our landlord.

At least, it became the landlord of one of our zColo sites. When Google agreed to buy 111 Eighth Avenue, Zayo's colocation business was operating in the building, alongside Equinix, Digital Realty and Telx. zColo was a small tenant in a big building, and the building's new owner was one of the companies reshaping our industry.

How zColo Started

I founded Zayo in 2007 to roll up fiber. Data centers were never the plan. But fiber and data centers are joined at the hip. The most valuable fiber in any city is the fiber that runs into the carrier hotels and data centers where networks meet. Bill LaPerch figured this out at AboveNet years before we did, and as I wrote in Bandwidth, he built his whole strategy around it: "we establish a fiber ring that hits the key carrier hotels and data centers."

On September 9, 2009, Zayo bought FiberNet Telecom Group for $104.1 million.Zayo's 10-K for fiscal 2011 gives $104.1 million. The acquisitions table in the fiscal 2019 10-K lists FiberNet at $96.6 million, apparently net of cash acquired. Along with the rest of the business came something unusual for a fiber company: space and interconnection inside carrier hotels. In the words of our 2011 annual report, "We formed our zColo business unit from a portion of the legacy FiberNet assets."

zColo started with sites in three of the most important carrier hotels in the New York area: 60 Hudson Street and 111 Eighth Avenue in Manhattan, and 165 Halsey Street in Newark. It also had sites in Los Angeles and Nashville. The crown jewel was 60 Hudson. We were, as our 2013 annual report put it, "the exclusive operator of the Meet-Me Room at 60 Hudson Street in New York City," a room where about two hundred global networks interconnected.The FiberNet origin is from Zayo's 10-K for fiscal 2011, and the Meet-Me Room description is from the fiscal 2013 10-K. Neither filing says when zColo was formally launched, and I haven't found a launch press release.

The Roll-Up

Zayo's first zColo acquisition was small. On December 31, 2011, we bought MarquisNet, which ran a single 28,000-square-foot data center in Las Vegas, for $13.6 million.

Six months later came the biggest deal in Zayo's history to that point. On July 2, 2012, we closed on AboveNet for $2.21 billion. I told that story in Bandwidth, and I won't repeat it here. For this book, what matters is that LaPerch's AboveNet came with exactly the fiber rings into carrier hotels and data centers that he had spent a decade building.

zColo kept growing in small bites. By August 2013, it had twenty-five data center and interconnection facilities in nineteen U.S. markets and about 175,000 billable square feet. Its revenue had grown from $43.3 million in fiscal 2012 to $63.4 million in fiscal 2013. That was still a sliver of Zayo, which booked $988.1 million of revenue that year.

Then the bites got bigger. We bought CoreNAP in Austin in May 2013 for $7.1 million and CoreXchange in Dallas in March 2014 for $17.2 million. On July 1, 2014, we paid $51.9 million for Colo Facilities Atlanta, better known as AtlantaNAP. The same day, we bought 96 percent of Neo Telecoms, which brought more than 300 route miles of fiber in Paris.

The biggest zColo deal was Latisys. When we announced it in January 2015, Data Center Knowledge noted that it was Zayo's first acquisition aimed mainly at data centers, after about thirty acquisitions since 2007. Latisys had eight data centers in five markets, Northern Virginia, Chicago, Denver, Orange County and London, with more than 185,000 square feet of billable space and 33 megawatts. It had more than $100 million of revenue, growing about 20 percent a year. Only Chicago overlapped with zColo's existing markets. We closed the deal on February 23, 2015, for $677.5 million, net of cash acquired.The announced price was $675 million. The fiscal 2015 10-K gives $677.5 million net of cash acquired; the fiscal 2019 10-K's acquisitions table gives $677.8 million.

Greg Friedman of zColo explained the logic to the trade press in the plainest terms: "Zayo's been historically a consolidator of infrastructure assets."

By mid-2015, zColo ran "45 datacenter facilities across 37 markets throughout the United States and France," and managed "more than 520,000 square feet of billable colocation space." Smaller deals followed: a Dallas data center from Stream at the end of 2015 for $16.6 million, a Santa Clara data center in 2016 for $11.3 million, KIO Networks' U.S. data centers in 2017 for $11.9 million, and a McLean, Virginia, data center in 2018.The dates and prices are from the acquisitions table in Zayo's 10-K for fiscal 2019: Stream Dallas Data Center, December 31, 2015, $16.6 million; Santa Clara Data Center, October 3, 2016, $11.3 million; KIO Networks US Data Centers, May 1, 2017, $11.9 million; McLean Data Center, April 4, 2018, no price shown. The same table lists forty-five deals through June 2019, totaling $6.6 billion, most of them fiber.

We also lost some. As I wrote in Bandwidth, Crown Castle bought Wilcon, a data center and fiber provider in Los Angeles, and I had tried more than once to buy it myself.

Was owning buildings a good idea for a fiber company? The case for it was simple. A meet-me room like 60 Hudson is where the networks are, and the networks are the customers for fiber. Owning the room puts you at the center of the action. The case against it was just as simple. Data centers eat capital, and a dollar spent on raised floor is a dollar not spent on fiber. And the biggest buyers of data-center space, the cloud companies and the REITs, were also some of the biggest buyers of fiber. Did we want to compete with our own customers for their tenants? Every fiber company that bought a building had to answer those questions, and few answered them the same way twice.

The Ceiling

By fiscal 2018, zColo's revenue had reached $290.7 million. In fiscal 2019, it fell to $269.9 million, "due to lower net installs." Zayo's consolidated revenue that year was $2.58 billion, so zColo was about a tenth of the company. Company-wide, Zayo operated about 1 million square feet of billable colocation space across fifty-one locations in the United States, Canada and Europe.

So zColo grew from a handful of carrier-hotel suites into a real business, and then it stopped growing.

The analysts had a view on what that business was worth. In March 2019, Nick Del Deo of MoffettNathanson took a close look at zColo. He gave Zayo credit for a strategy he summed up as "acquire diamonds in the rough, clean them up, and be left with polished gems." Then he did the math. At 12 times reported EBITDA for zColo, the market was valuing the rest of Zayo at about 9.2 times. At 13 times his own EBITDA estimate for zColo, the core fiber business was implied at about 11.75 times. His verdict on zColo was polite and hard to argue with: "its modest contribution to the overall business means it wouldn't deliver much of a punch."Nick Del Deo and Sydney Marks, "Zayo: Plan A, Version 2.0… A Closer Look at zColo," MoffettNathanson, March 4, 2019. The note came after Zayo dropped a plan to split itself in two and positioned zColo for a sale or a combination. Del Deo gave lower odds than most observers to a sale of Zayo itself. The multiples are from the note's exhibits as summarized on its first pages; I haven't read the full PDF. Del Deo later hosted me for a meeting with his clients, in February 2025, after Bandwidth came out.

A year later, in Zayo's last quarter as a public company, Del Deo reported that the fiber business had grown revenue 7.0 percent while zColo's had fallen 5.3 percent. He called them "poor zColo results." Twelve or thirteen times EBITDA is a respectable price for a business that is shrinking. It is not much of an argument for keeping one.

The Sale

The end of zColo came in two acts, and the first was a bit of theater.

In 2019, while Zayo was negotiating its own sale, I wrote in Bandwidth, "more as a ruse than a contingency plan, we launched a process to sell our data center business." The point was to keep the bidders for Zayo on their toes. The full story of that sale is in Bandwidth. Both bidding groups were serious, capable investors. The data-center process was a sideshow.

It was a sideshow with a full cast, though. Zayo's merger proxy records that on December 21, 2018, a Canadian-headquartered alternative asset manager, which the filing calls "Party A," called me "to discuss a potential purchase of the Company's colocation business, rather than all of the Company." When one of the groups bidding for all of Zayo made its proposal in February 2019, the proposal was contingent on the proceeds of any zColo sale "remaining with the Company." On March 6, Zayo began to solicit interest formally. J.P. Morgan contacted about fifty parties, and twenty-one signed confidentiality agreements. Four bids came in on April 10. The highest, from Party A, was "generally in line with management's valuation expectations." A ruse can still tell you what something is worth.

The second act was real. Digital Colony and EQT took Zayo private in March 2020, as we'll see in the next chapter, and the new owners decided to sell zColo for real. A year earlier, the trade press had reported that the sale could fetch up to $1 billion.

On September 29, 2020, DataBank agreed to buy zColo. DataBank was majority-owned by Digital Colony, which was also one of Zayo's two new owners, alongside EQT. zColo's forty-four facilities would join DataBank's twenty, taking DataBank from nine metros to twenty-nine. Its revenue would grow from about $180 million to about $450 million. Zayo would become DataBank's largest customer. DataBank's president, Kevin Ooley, said what any buyer would want to say about a carve-out: "The [zColo] business has a healthy EBITDA margin."

Neither side disclosed a price. Colony Capital's annual report for 2020 eventually did: "In December 2020, our DataBank subsidiary closed on its acquisition of zColo, the colocation assets of Zayo Group Holdings, Inc. ("Zayo"), consisting of 39 data centers in the U.S and U.K., for approximately $1.2 billion." Outside co-investors put up $0.5 billion of the financing. Zayo, the filing said, "will continue to be an anchor tenant within the zColo facilities."Baxtel reported the price as $1.4 billion on September 30, 2020, with no attribution, and that figure has been widely repeated. Colony's 10-K is the primary source, and its $1.2 billion covers only the thirty-nine U.S. and U.K. data centers. The price of the five French data centers hasn't been disclosed. Colony's own balance-sheet investment was $188 million, falling to about $145 million after a February 2021 capital raise. Data Center Frontier likewise did not disclose a price; it reported that zColo added forty-four data centers in twenty-three U.S. and European markets and that Colony put $145 million from its balance sheet into the deal. Advisers, per search snippets only: DH Capital and Jones Day for DataBank; J.P. Morgan and Skadden for zColo.

Data Center Frontier's report on the deal quoted me: "This agreement allows both parties to focus on their core strengths." That was the whole logic. Zayo did fiber. DataBank did buildings.

The U.S. and U.K. sites closed on December 15, 2020. The five French sites closed separately, on February 3, 2021. "We're excited to finally have all elements of the zColo transaction complete," DataBank's CEO, Raul Martynek, said.

As I wrote in Bandwidth, after Zayo was taken private, the new investors sold its data-center business, and $600 million of the proceeds were distributed to investors and lowered net indebtedness.

Let's do the arithmetic, with the caveat that the pieces don't line up perfectly. FiberNet cost $104.1 million in 2009, and only part of it became zColo. The data-center acquisitions that followed, from MarquisNet to KIO, added up to roughly $800 million, most of it Latisys. Eleven years later, the U.S. and U.K. sites alone sold for about $1.2 billion, and the French sites went separately. Is that a good return? I can't say from the filings alone, because they don't show what we spent building out and running the sites over those eleven years. For a business that was never the plan, it ended up worth more than a billion dollars to somebody else. Nobody would write a book about it, though, which is why it gets a section in this one and not a chapter.

A Handful of Buyers#

By 2016, the data-center business looked nothing like it had in 2006.

Ten years earlier, the industry's customers were thousands of companies renting cages and racks. By 2016, AWS alone was running at about $10 billion of revenue a year, and every dollar of it ran in a data center that Amazon either built or leased. Google, Microsoft and Facebook had become the largest builders and the largest tenants in the industry at the same time. Facebook had published its building designs and told the world the building wasn't its competitive advantage. Google owned one of the most important carrier hotels in New York. Loudoun County had 10 million square feet and was still building every day.

In Bandwidth, I told how the fiber business ended up with a handful of national platforms. The data-center business was heading the same way from the opposite direction. The demand had consolidated first. A handful of cloud buyers now controlled most of the new demand, and each of them could build without you.

That shifted power from the landlords to the tenants. In the Gold Rush, the landlord set the price and the dot-com paid it. By 2016, the tenant knew the landlord's cost to the dollar, had a construction team of its own and could take its business to Quincy or Prineville if the price was wrong.

The rents show it. According to Cushman & Wakefield, wholesale space in Northern Virginia leased for about $143 per kilowatt per month in 2017. By 2020, it was $103, down 28 percent, in the busiest data-center market in the world. The analysts saw the same thing from the other side. In July 2019, after two slow quarters of hyperscale leasing, MoffettNathanson's Del Deo concluded that the cloud's demand for space was intact, "even if it manifests in a lumpy fashion." When five customers make up the market, demand arrives in lumps, and every landlord waits for the next one.

The power bill showed the same concentration, and it gave the cloud a better report card. Remember EPA's 2007 forecast: more than 100 billion kilowatt-hours by 2011. It never came. In 2016, the Lawrence Berkeley National Laboratory found that U.S. data centers had used about 70 billion kilowatt-hours in 2014, about 1.8 percent of the nation's electricity. Use had grown nearly 90 percent from 2000 to 2005, 24 percent from 2005 to 2010 and only about 4 percent from 2010 to 2014.Arman Shehabi et al., United States Data Center Energy Usage Report (LBNL, June 2016). It cites a 2011 update of EPA's method that put 2008 use at 69 billion kilowatt-hours, also 1.8 percent. A later LBNL report, in 2024, revised the level for 2014–2016 to about 60 billion kilowatt-hours using a different method; the two series shouldn't be spliced. The 2016 report projected about 73 billion kilowatt-hours for 2020. Why? The report gave the answer in one line: "Nearly all server shipment growth since 2010 occurred in servers destined for large hyperscale data centers." The computing kept growing. It just moved into the giants' efficient buildings. Had the industry stayed at its 2010 efficiency, the report estimated, it would have needed more than 600 billion additional kilowatt-hours over the decade.

So two forecasts missed in this chapter, in opposite directions. Loudoun's square-footage forecasts came in too low. Washington's power forecast came in too high. Odlyzko would have recognized the second miss. The question for the end of this book is which kind of miss the AI forecasts turn out to be.

None of this made the landlords' business bad. Data-center stocks had come through the crisis better than almost anyone expected, and the cloud companies were the best tenants anyone had ever had. But the landlords were no longer selling to a crowd. They were selling to a club, and the club's members could build their own clubhouse.

How do landlords respond when their customers consolidate? The same way the carriers responded in Bandwidth: they consolidate too. In May 2015, Equinix agreed to buy TelecityGroup in Europe. Two months later, Digital Realty agreed to buy Telx. And a tower investor named Marc Ganzi was about to discover data centers.

The Reconsolidation was underway.

On the morning of Monday, November 15, 2021, two press releases crossed the wire within hours of each other.

In the first, KKR and Global Infrastructure Partners agreed to buy CyrusOne, the data-center company Cincinnati Bell had spun off nearly nine years earlier, for about $15 billion including debt. Data Center Frontier called it "the largest M&A deal in the history of the data center sector." In the second, American Tower, one of the big three cell-tower companies, agreed to buy CoreSite, the Denver data-center REIT, for about $10.4 billion including debt.

Five months earlier, Blackstone had agreed to buy QTS for about $10 billion. Six months later, Marc Ganzi's DigitalBridge would agree to buy Switch for about $11 billion. Four data-center companies, about $46 billion of value, in eleven months.The four are QTS (about $10 billion, announced June 7, 2021), CyrusOne (about $15 billion, November 15, 2021), CoreSite (about $10.4 billion, November 15, 2021) and Switch (about $11 billion, May 11, 2022), all including debt. That's about $46 billion over eleven months of announcements. The QTS and CoreSite figures follow the companies' SEC filings; MoffettNathanson's same-day estimates of enterprise value were lower for QTS (about $8.5 billion) and the same for CoreSite, and on its figures the four come to about $45 billion. See notes 13a and 16. Every deal in this chapter is listed with its date, price and source in Table 7.1 at the end of the chapter.

In Bandwidth, I opened the chapter on the fiber industry's Reconsolidation with a line from an analyst in January 2002: "What you have been seeing is the gradual reconsolidation of an industry that is highly capital intensive and that's perhaps best served by a relatively small number of large companies." It took the fiber business more than a decade to prove him right.

The data-center business did it in about seven years.

At the end of Chapter VI, the demand had consolidated. A handful of cloud companies controlled most of the new demand, and each of them could build without a landlord. This chapter is the supply side's answer. Between 2015 and 2022, the two public giants, Digital Realty and Equinix, bought their largest rivals on both sides of the Atlantic. The phone companies that had bought data centers at the top of the last cycle sold them. And private capital, the infrastructure funds, Blackstone, KKR and a former tower developer named Marc Ganzi, bought nearly everything else, including several of the REITs that had gone public in Chapter VI.

The phoenix had grown up. Now the big birds were eating the small ones.

Digital Realty—Bill Stein#

When we last left Telx in Chapter VI, it was a colocation tenant at 111 Eighth Avenue, renting space in a building that Google had just bought. Telx was in Equinix's business, retail colocation: renting cages and racks to networks and enterprises, and charging them to connect to one another. It was the opposite of Digital Realty's business of leasing big, powered buildings to a few big tenants.

On July 14, 2015, Digital Realty agreed to buy Telx from its private-equity owners, ABRY Partners and Berkshire Partners, for $1.886 billion in cash. Telx had twenty data centers. Bill Stein of Digital Realty explained why the fit was so natural: "The fact that more than half of Telx's 20 facilities are run out of Digital Realty properties." Digital Realty was buying its own tenant.Digital Realty backed the purchase with a $1.85 billion unsecured 364-day bridge-loan commitment and expected it to add to 2016 earnings. Digital Realty's 8-K dates the merger agreement July 13, 2015, and the release names Stein as Digital Realty's chief executive. The deal closed on October 9, 2015.

Stein called it "transformative," and he was right. With Telx, Digital Realty was in Equinix's business. The wholesale landlord could now sell a cloud company a building and sell the networks inside it the connections between them.

DuPont Fabros

Two years later, Digital Realty went after the company that had followed it to market.

When we last left DuPont Fabros, it had gone public at $21 a share in October 2007, fallen to $1.93 in the credit crisis, and come back as the cloud companies filled its buildings. Facebook had been one of its early tenants in Ashburn.

The merger proxy tells how it happened. On March 15, 2017, Stein called DuPont Fabros's CEO, Christopher Eldredge, and set up a lunch. Over that lunch in Ashburn, on March 24, Stein offered 0.51 to 0.53 Digital Realty shares for each DuPont Fabros share, worth $52.50 to $55.00, all in stock. Eldredge said no. Stein came back in April, and DuPont Fabros said that wasn't enough either and hired Goldman Sachs. Digital Realty raised its ratio twice in May, to 0.545, and the two sides signed an exclusivity agreement.

Then it nearly fell apart. On May 31, DuPont Fabros's chairman, Lammot J. du Pont, told Stein that more negotiation "would not be productive." On Sunday, June 4, DuPont Fabros "turned off access to the virtual data site" and asked Digital Realty to return or destroy its documents. The next day Digital Realty sent a new draft, and the deal was back on. The last issue was board seats. Both boards approved on June 8.

On June 9, 2017, Digital Realty announced an all-stock merger with DuPont Fabros valued at about $7.6 billion, including $1.6 billion of assumed debt. DuPont Fabros had twelve operating data centers and six more under development. Eldredge said, "We are excited to deliver this compelling transaction to our shareholders." Stein called it "strategic and complementary."The exchange ratio was fixed at 0.545 Digital Realty shares per DuPont Fabros share. The release gave no implied per-share price or premium. The termination fee was $150 million payable by DuPont Fabros or $300 million payable by Digital Realty in certain circumstances, and the deal had no financing condition. The merger closed on September 14, 2017, at an enterprise value of about $7.8 billion.

So the first data-center REIT bought the second. The wholesale pioneer that had mothballed a $270 million Santa Clara project in 2008 now commanded a $7.6 billion enterprise value. Its big cloud tenants became Digital Realty's tenants.

The Jilted Bride

The third deal had a backstory.

By the spring of 2015, Interxion, a European data-center operator, had agreed to an all-stock merger with a rival, TelecityGroup, valued at $2.2 billion. Then Equinix came along with a better offer for Telecity, as we'll see in the next section. Telecity dropped Interxion.

Interxion and Digital Realty spent the next four years circling each other. According to Digital Realty's exchange-offer prospectus, they first met in March 2016, and talks broke off and restarted four times, through data rooms and draft purchase agreements, before the final round began in August 2019. On October 10, 2019, an industry publication reported that Interxion was drawing interest, and the leak didn't scare up a rival. Only after signing did Digital Realty learn that, over four years, Interxion had held "extensive contacts" with "eight counterparties," including "a sovereign investment fund." None of them bid.

On October 29, 2019, Digital Realty agreed to combine with Interxion in a deal worth about $8.4 billion, including net debt. Interxion's shareholders would get Digital Realty stock worth about $93.48 a share and own about 20 percent of the combined company. Interxion brought fifty-three data centers in eleven European countries. Together, the companies would reach about 70 percent of Europe's GDP. Interxion's CEO, David Ruberg, said the goal was to "build the world's preeminent data center provider."

The deal closed on March 12, 2020, three days after Zayo became a private company.The exchange ratio was fixed at 0.7067 Digital Realty shares per Interxion share, based on Digital Realty's $132.28 close on October 28, 2019. The exchange offer expired at 12:01 a.m. on March 12, 2020, with about 92.3 percent of shares tendered. Digital Realty's prospectus puts the 0.7067 ratio at about an 18 percent premium to Interxion's unaffected close of $78.14 on October 9, 2019, the day before the press report. Because fewer than 95 percent were tendered, the remaining holders were paid through a Dutch liquidation distribution, generally subject to 15 percent Dutch withholding tax. The merger release said Ruberg would run Digital Realty's European business, branded "Interxion, a Digital Realty company," for about a year after closing.

Add it up. In under five years, Digital Realty spent about $1.9 billion in cash on Telx and issued stock for DuPont Fabros and Interxion in deals worth about $16 billion more, including debt. It went from a wholesale landlord in the United States to a global company that did wholesale and retail, in America and Europe. Did it overpay? The stock-for-stock structure meant Digital Realty's shareholders shared the risk with the sellers, which is the honest way to do a big deal. It also meant Digital Realty didn't have to answer that question in cash.

Equinix—Steve Smith and Charles Meyers#

When we last left Steve Smith in Chapter V, he had taken over Equinix in 2007 and spent the Phoenix years expanding it, from IXEurope to Switch and Data. Steve is a friend of mine. He ran Equinix for about a decade and later took my old chair at Zayo. I'll tell what the record shows here and leave the color to him.

Telecity

On May 29, 2015, Equinix agreed to buy TelecityGroup for £2.35 billion, about $3.6 billion, in cash and stock. The price was a 27.3 percent premium to Telecity's share price before the talks were disclosed.The terms were 572.5 pence in cash plus 0.0327 of a new Equinix share for each Telecity share, about 1,145 pence in all. The premium is measured against Telecity's May 6, 2015 close. Telecity's holders would own 10.1 percent of the combined company, and Telecity's executive chairman, John Hughes, joined Equinix's board. Reuters put the price at £2.35 billion; The Register reported £2.3 billion. Telecity's board dropped its $2.2 billion deal with Interxion to take it.

Smith explained the logic in one sentence: "Their customer base and footprint are more complementary to where we needed to go next." Equinix had roughly a hundred data centers at the time. Telecity added thirty-nine in Europe.

By the time the deal closed, on January 15, 2016, Equinix's share price had moved enough that the value had grown to about $3.8 billion. Equinix gained more than forty data centers, roughly doubled its European capacity and entered seven new countries: Bulgaria, Finland, Ireland, Italy, Poland, Sweden and Turkey. Together with another acquisition, Bit-isle, Telecity took Equinix from 105 data centers in thirty-three metros to 145 in forty.The closing value was about $1.7 billion in cash plus about 6.8 million Equinix shares worth about $2.1 billion, or about £2.6 billion in all. The 145-data-center count is after divestitures the European Commission required. Eight of the divested sites went to Digital Realty for about $874 million in July 2016. "Today is a very important day for Equinix," Smith said.

Verizon's Data Centers

The next deal reversed one of the bigger bets of the Phoenix years.

As we saw in Chapter V, Verizon had bought Terremark in 2011 for about $1.4 billion of equity, to build a cloud business on top of Terremark's data centers. Verizon already owned the data centers it had inherited from MCI. By 2016, Verizon had decided to focus on wireless and its FiOS fiber-to-the-home business.

At the time of that deal, Terremark ran thirteen data centers and Verizon more than 220. Verizon's plan was to marry Terremark's cloud and hosting skills to Verizon's enterprise customers. It was the same plan the carriers had pitched in the Gold Rush, when Qwest built CyberCenters and Level 3 built Gateways. The phone company would own the network and the buildings, and sell the customer both.

In December 2016, Equinix agreed to buy twenty-nine of Verizon's data centers in the United States and Latin America for $3.6 billion in cash. The sites had 2.4 million square feet and about nine hundred enterprise and government customers in fifteen metros. Verizon kept its managed-hosting and cloud businesses. It said the sale would "focus resources in areas that will help drive digital transformation for enterprise customers." Smith said the deal "significantly increases the size and sophistication of our enterprise and government ecosystems."The Register counted twenty-nine data centers, twenty-four of them customer-facing. Washington Technology's headline at closing called it a "$3B deal" but its text gave the estimated $3.6 billion. I haven't read Equinix's own closing release, so the exact closing date is unconfirmed.

The deal closed around May 1, 2017. Seven of the data centers were in Virginia, and four of them supported government interconnection. About six hundred of the customers were new to Equinix. After the deal, Equinix had 175 data centers in forty-four markets and 17 million gross square feet.

Was Verizon right to sell? Verizon had paid about $1.4 billion for Terremark's equity in 2011 and sold twenty-nine data centers for $3.6 billion five years later. Those aren't the same assets, so the comparison flatters Verizon. But the direction was clear. The phone company wanted out, and the company that knew what to do with the buildings wanted in.

xScale

By 2019, Equinix had a problem that came straight out of Chapter VI. Its business was retail: thousands of customers renting cages and paying for connections. Its biggest growth opportunity was the cloud companies, which wanted whole buildings at wholesale prices. Should Equinix build those on its own balance sheet, where lower wholesale returns would sit next to its higher retail ones?

The answer was to borrow someone else's balance sheet. On July 1, 2019, Equinix announced a joint venture with GIC, Singapore's sovereign wealth fund, worth more than $1 billion, to build six hyperscale data centers in Frankfurt, London, Amsterdam and Paris, with about 155 megawatts of capacity. GIC would own 80 percent and Equinix 20 percent. Equinix contributed two sites, LD10 in London and PA8 in Paris. The target customers were Google, Microsoft, Oracle, AWS and Alibaba Cloud. Equinix called it xScale.

Charles Meyers announced it for Equinix: "we are tremendously excited to announce the formation of our first xScale data centers joint venture."Meyers was Equinix's president and CEO. Smith had resigned early in 2018, and Meyers, then a senior Equinix executive, was named CEO on September 12, 2018, after Peter Van Camp's stint as interim CEO. The agreement was signed on June 27, 2019, and Equinix and GIC completed the joint venture's formation on October 9, 2019. Equinix insisted it wasn't entering the wholesale business broadly. Fitch upgraded Equinix's credit rating to BBB the same day.

The retail king had found a way to serve the hyperscalers without carrying the buildings on its own balance sheet alone. Most of the money came from a sovereign fund. That kind of money was about to show up everywhere in this business.

Steve Smith, as I wrote in Bandwidth, went on to replace me as Zayo's CEO in 2020, six months after Zayo went private. Which brings us to the man who bought Zayo.

DigitalBridge—Marc Ganzi#

In the mid-2010s, I met Marc Ganzi for the first time at a small investor conference at the Deer Valley Resort in Utah. As I wrote in Bandwidth, when I introduced myself, he asked with a beaming smile, "You know what investors call me?" I shrugged. "They call me the Dan Caruso of Towers." He paused, smirked and added, "And you are known as the Marc Ganzi of Fiber."

Ganzi had co-founded Global Tower Partners in 2003 and sold it to American Tower in 2013 for $4.8 billion. That same year, he and Ben Jenkins, a former Blackstone executive, co-founded Digital Bridge to focus, as I described it in Bandwidth, on the broader communications-infrastructure sector.

Ganzi and I were, as I wrote in Bandwidth, cut from the same control-freak cloth. We both needed to be the top dog, and I couldn't imagine a scenario where I'd work for him. His record in data centers follows a playbook any fiber roll-up veteran would recognize: buy a platform, keep a strong operator in charge and grow it with more acquisitions.

The Platforms

Digital Bridge's first data-center platform was DataBank, a Dallas company it bought from Avista Capital Partners in July 2016. The price wasn't disclosed.Terms weren't disclosed. Ganzi told Reuters the deal was "a great launching pad to do more investments in data centers."

Its second was Vantage, the Santa Clara wholesale developer Silver Lake had launched in 2010. In March 2017, Digital Bridge, PSP Investments and TIAA Investments bought it, again for an undisclosed price. Vantage had 51 megawatts built in Santa Clara, room to grow to 93, and a 6-megawatt campus in Quincy, Washington, the potato town of Chapter VI. Sureel Choksi stayed as CEO. The new chairman was Michael Foust, Digital Realty's co-founder, who had been its CEO when it went public as the first data-center REIT.

Ganzi summed up his thesis in one line: "The data center space is actually in the early innings."

Zayo

Then Ganzi came for Zayo.

The full story of the Zayo sale is in Bandwidth, and I won't retell it here. The short version: in 2018, with activist investors Starboard Value and Sachem Head pressing the company and its stock well off its highs, Zayo drew interest from two groups of the largest infrastructure investors in the world. One was a consortium of Blackstone, I Squared, KKR and Stonepeak, partnered with Zayo's biggest private investors, Charlesbank and GTCR. The other was Ganzi's Digital Colony, teamed with EQT. Both were serious, well-financed bidders.

The merger proxy Zayo filed in June 2019 gives the dates, in the flat language lawyers use. In mid-July 2018, at Ganzi's suggestion, he and I met in Aspen. We met again in Boulder in early August. On October 19, Ganzi called about buying Zayo together with its management. On November 6, just before a Zayo board meeting, he said he was arranging a "fully financed proposal" at $41 to $42.50 a share. Two days later, Zayo reported earnings and announced a plan to separate the company in two, and the stock fell from $30.38 to $22.56.

The bids that followed never got near Ganzi's talk. The proxy doesn't name the bidders; it calls the first group Consortium A and the Digital Colony and EQT group Consortium B. Consortium A bid $33.00 in November, raised to $35.00 in December, cut to $31.50 after due diligence in February 2019 and offered $32.00 as its "best and final." The board's answer was that "the Company would not be transacting at $31.50 per share and that any leaks would not be tolerated." Consortium B had put a written range of $36 to $38 in front of the board in December 2018. Its final price was $35.00.Zayo Group Holdings, definitive proxy statement, June 26, 2019, "Background of the Merger." The proxy's parties are anonymized; the names in the text are from Bandwidth, which calls the two groups "Wind" and "Sun." Consortium A's bids were $33.00 (November 16, 2018), $34.50 (November 19), $35.00 (December 7), $31.50 (February 4, 2019) and $32.00 (February 20). Ben Jenkins of Digital Colony sent Consortium B's $36–$38 range on December 16, 2018, during a board meeting.

On May 8, 2019, Zayo agreed to be acquired by affiliates of Digital Colony Partners and the EQT Infrastructure IV fund for $35 a share in cash, about $14.3 billion including $5.9 billion of net debt. In the press release, I said, "Digital Colony and EQT share our vision that Zayo's Fiber Fuels Global Innovation." In the memo to our employees that morning, I wrote, "I am incredibly proud of the company we have built and am excited about Zayo's future."Bloomberg put the equity value at $8.2 billion and the premium at 14.3 percent to the prior day's close. Zayo's release measured the premium as 32 percent against the six-month volume-weighted average price of $26.44. Zayo's shares had peaked at $39.35 the previous July and fallen as low as $20.46 before the company announced a strategic review in March 2019. Bloomberg called it the largest transaction EQT had announced in its twenty-five-year history.

The deal closed on March 9, 2020. As I wrote in Bandwidth, it became the second-largest leveraged buyout since the 2008 financial crisis. Six months later, I left, and Steve Smith took over.

zColo

Zayo's own data centers, zColo, had been for sale for most of that time. The proxy records the process. On December 21, 2018, a Canadian-headquartered alternative asset manager, called Party A in the filing, phoned me "to discuss a potential purchase of the Company's colocation business, rather than all of the Company." On March 6, 2019, Zayo "began to formally solicit interest." J.P. Morgan contacted about fifty parties, and twenty-one signed confidentiality agreements. On April 10, four bids came in. The highest was Party A's, "generally in line with management's valuation expectations." Consortium B's $35 offer was conditioned on the proceeds of any zColo sale staying inside Zayo.Same proxy. J.P. Morgan's count later rose to fifty-two parties, and the twenty-one confidentiality agreements were signed between March 19 and April 8, 2019. The proxy doesn't name Party A, and this book doesn't either. MoffettNathanson's views are from Del Deo and Sydney Marks, "Zayo: Plan A, Version 2.0… A Closer Look at zColo," March 4, 2019, and Del Deo, "Zayo: Q2 F2020 Earnings: Good Night, and Good Luck," February 4, 2020. Data Center Frontier did not disclose the zColo price; Table 7.1 uses Data Center Knowledge's figure of about $1.2 billion for thirty-nine sites; other reports were higher.

In Bandwidth, I called the data-center sale "more as a ruse than a contingency plan," a bit of noise to keep both bidders on their toes. The filing shows the noise was run as a real auction, with a real price.

The outside view of zColo was cooler. In March 2019, MoffettNathanson's Nick Del Deo valued it at about twelve to thirteen times EBITDA and concluded that its "modest contribution to the overall business means it wouldn't deliver much of a punch." In Zayo's last quarter as a public company, he reported, zColo's revenue fell 5.3 percent from a year earlier.

zColo didn't sell while Zayo was public. As we saw in Chapter VI, the new owners sold it to DataBank in September 2020. Digital Colony owned part of the seller and most of the buyer, and Colony put $145 million from its own balance sheet into the deal to keep its 20 percent stake in DataBank. Data Center Frontier reported that zColo added forty-four data centers in twenty-three markets, taking DataBank from twenty data centers to sixty-four. Data Center Frontier quoted me on the deal: "This agreement allows both parties to focus on their core strengths." About $600 million of the proceeds went to Zayo's investors.

Colony

The most surprising part of Ganzi's arc happened at the same time.

In July 2019, two months after the Zayo agreement, Colony Capital, the real-estate investment firm founded by Tom Barrack, agreed to buy Digital Bridge Holdings for $325 million. The plan was for Ganzi to become Colony's CEO in eighteen to twenty-four months, with Barrack returning to executive chairman. Colony managed about $43 billion of assets, and the deal would take it to about $60 billion. Barrack called Ganzi "an exceptional choice."

Ganzi became CEO of Colony in 2020. Then he did something few CEOs of a big, diversified company would try: he sold off most of what wasn't digital. Colony sold about $2.7 billion of legacy assets to Fortress. On June 22, 2021, Colony Capital became DigitalBridge Group, trading on the New York Stock Exchange as DBRG. Ganzi's goal was to rotate $100 billion of assets under management from diversified real estate to digital infrastructure. Pro forma, the firm was 80 percent of the way there.

"We are poised to build, operate, and finance a new era in connectivity," he said.

Think about the round trip. In 2019, Ganzi sold his company to Colony for $325 million. Two years later, Colony had taken his company's name, his strategy and his leadership. The buyer had become the bought.

Switch

Ganzi's biggest data-center deal came last.

When we last left Rob Roy in Chapter V, he was a Las Vegas real-estate kid who had bought a small former Enron Broadband building, wired to many carriers, for about $930,000 at a private bankruptcy sale in late 2002.The details of Switch's Enron origin are murky. Court records cited by Data Center Knowledge name the buyer as "Colo Gateways," not Switch or Roy, and put the building at 12,000 square feet, so the Enron facility was not the later SuperNAP. Switch's own 2017 IPO prospectus, in the portion I read, doesn't mention Enron. See Chapter V. Twenty years later, his company, Switch, was one of the largest listed U.S. data-center companies after Digital Realty and Equinix, according to Bloomberg. It had opened its fifth "Prime" campus, in Texas, in 2021 and planned more than 11 million square feet of additional space by 2030.

Switch's proxy shows a real contest. Its bankers contacted sixteen potential buyers, and nine signed confidentiality agreements. In the first round, in March 2022, DigitalBridge bid $34.00 a share and a rival the proxy calls Bidder A bid $32.50. Then the buyers hit a snag that only a Phoenix company could produce: some of Switch's Las Vegas sites sat on land owned in part by one of its own directors, and the bidders needed the leases changed to finance the deal. The owners refused, and offered to sell the land instead, for $300 million. In May, Bidder A dropped to $29.00, citing the cost of financing, and then came back at $32.50 on the condition that Switch sign that day. DigitalBridge and IFM held at $34.25, "best and final," a price that also had to cover the land and a $75 million buyout of a tax agreement.

On May 11, 2022, DigitalBridge and IFM Investors agreed to buy Switch for $34.25 a share in cash, about $8.4 billion for the equity and about $11 billion including debt. The price was a 19 percent premium to Switch's share price before Bloomberg reported a sale process. Bloomberg's sources named the losing bidder as Brookfield Asset Management.The premium is measured against Switch's March 18, 2022 close. The proxy calls the rival "Bidder A"; it cites the May 3 and May 9, 2022 press reports that named Brookfield. The Las Vegas land purchase arose because director Tom Thomas had an ownership interest in the sites; he recused himself. Roy's share of the tax-agreement buyout was about $11.7 million. Switch's special committee approved the deal unanimously. Goldman Sachs and Morgan Stanley advised Switch; RBC led for the buyers, with TD as co-adviser and debt lead.

Roy pointed to "our advanced datacenter infrastructure" and the "significant expansion capacity in our land bank." According to the proxy, Roy and Thomas Morton rolled about $500 million of their equity into the new company. The deal closed in December 2022.

From $930,000 for an Enron building to about $11 billion, in twenty years. That's the kind of Phoenix story Bandwidth was built on.

The tower guy who called himself the Dan Caruso of Towers now ran a firm whose platforms included DataBank, Vantage and Switch. The record shows he moved into data centers before most of the big infrastructure investors did, and through 2022, "early innings" looked right. The record also shows a firm that had bet its whole identity on one sector, which is a fine strategy until the sector stumbles.

The Take-Private Wave—Blackstone, KKR, GIP and American Tower#

In Chapter VI, four wholesale developers went public as REITs: DuPont Fabros, CoreSite, CyrusOne and QTS. By the end of 2022, none of them was still public. Digital Realty bought DuPont Fabros. Blackstone took QTS private, KKR and GIP took CyrusOne, and American Tower, a public company itself, bought CoreSite. DigitalBridge took Switch private too.

Blackstone and QTS

QTS's proxy opens its account of the sale with a sentence that could stand as this chapter's thesis: "Building and managing data centers is an inherently capital-intensive business." QTS expected to spend about $1 billion on construction in 2021. It had raised about $860 million of new equity since January 2020 and had sold stock in "nearly 70% of the months." It was on a treadmill, and it knew it. In 2018, it and its bankers "contacted approximately 30 parties" looking for capital. Only one deal came of it, a joint venture on a single building in Manassas.

Blackstone first met QTS in December 2018, as a "capital provider." The relationship took two and a half years to turn into a bid. In September 2020, QTS's Chad Williams and Jeff Berson met Blackstone in Connecticut. A private-equity firm, an infrastructure fund and two pension funds looked and passed on price. After a dinner in New York on May 6, 2021, Blackstone offered $75.00 a share in writing. It went to $75.75. QTS's board asked for $78.00, and on May 14 Blackstone agreed. Blackstone also agreed to pay a reverse termination fee of 11 percent of QTS's equity, "more than $800 million," if it walked. During a forty-five-day window to shop for a better offer, the bankers contacted seven more parties. None bid.

On June 7, 2021, QTS announced that it would be acquired by Blackstone funds for $78 a share in cash, about $10 billion including debt. The price was a 21 percent premium to QTS's closing price the previous Friday, $64.49.The sources measure the deal differently. QTS's proxy puts the premium at 24.5 percent to the May 14, 2021 close of $62.64, the day Blackstone agreed to $78. The proxy also gives QTS's termination fees as percentages of both equity value and enterprise value, and those ratios imply an enterprise value of about $10 billion, matching Blackstone's release. MoffettNathanson's Del Deo and Michael Srour put the equity value at about $6.0 billion and the enterprise value at about $8.5 billion. I have used the SEC filing. The proxy also says Williams's affiliates were expected to receive about $140 million to $150 million in cash. QTS had gone public in 2013 at $21. Eight years later, Chad Williams's shareholders were being paid about 3.7 times the IPO price, before dividends.

The buyers were Blackstone Infrastructure Partners and Blackstone Real Estate Income Trust, known as BREIT, along with other Blackstone vehicles with what the firm called perpetual capital: money that doesn't have to be returned to investors on a fixed schedule. "We are focused on investing in assets that are benefitting from strong, secular tailwinds," said Tyler Henritze of Blackstone Real Estate. Greg Blank of Blackstone Infrastructure Partners said Blackstone was "delighted to back QTS and its world-class management team as they continue to scale the company."The premium was 24 percent against QTS's ninety-day volume-weighted average price. QTS would be jointly owned by Blackstone Infrastructure Partners and BREIT. Blank said QTS aligned with "one of Blackstone's highest conviction themes – data proliferation."

That phrase, perpetual capital, is the key to this whole chapter. A traditional private-equity fund buys a company, improves it and typically sells it within several years. A perpetual-capital vehicle can hold an asset indefinitely, the way a pension fund holds a bond. For that kind of owner, a data center leased to a cloud giant on a long lease looks less like a real-estate bet and more like a bond with upside. And it solves QTS's problem. A company that had to sell stock most months to keep building could now draw on an owner that never had to sell anything.

Not everyone thought QTS was worth it. On the day of the deal, Del Deo raised his target to the $78 deal price and admitted, "We've long struggled with QTS's valuation." He admired the company's "scrappy mindset" but wrote that its "bottom line results have never quite matched its qualitative appeal."

KKR, GIP and CyrusOne

The CyrusOne deal was the biggest of the wave. KKR and GIP agreed to pay $90.50 a share in cash, about $15 billion including debt, a premium of about 25 percent to CyrusOne's share price before rumors of a sale.The 25 percent premium is measured against CyrusOne's unaffected close on September 27, 2021, before reports of a sale. Data Center Frontier reported a premium of 2.5 percent, measured against the Friday close just before the announcement, after the rumors had already lifted the stock. CyrusOne had gone public in 2013 at $19. The buyout price was almost five times that, before dividends.

The deal also came at a moment of turmoil at the top. CyrusOne had more than forty data centers and about 190 of the Fortune 1000 as customers. But its CEO, Bruce Duncan, had left in July 2021 after about a year in the job, and co-founder Dave Ferdman had returned as interim CEO, the company's fourth chief executive in about eighteen months. "This transaction is a testament to the tremendous work by the entire CyrusOne team," Ferdman said.

Was the churn a reason to sell? CyrusOne's proxy suggests it was. CyrusOne had tried to sell itself once before, in 2019. Its bankers contacted more than fifteen parties, and KKR's group said any bid "would likely be lower than the then-current trading price." That process died. Duncan left on July 29, 2021. Thirteen days later, on August 11, the first bid arrived, at $80 a share.

What followed was a crowded auction. After the press reported the sale process in late September, fourteen potential buyers called. Bids came in at $82, $82.50 and $82 to $85. One bidder cut its price and lost its access to the data room. KKR's original partner dropped out in October, and KKR asked to team up with GIP instead. On November 14, 2021, the last day, KKR and GIP bid $90.50 a share against a rival's $86.00. The board took the higher number that evening.

KKR was a newcomer to data centers through a platform called Global Technical Realty. GIP managed more than $79 billion. The deal closed in late March 2022.CyrusOne Inc., definitive proxy statement, December 30, 2021, "Background of the Merger." The 2019 approach that started the failed process was an all-stock offer of about $64.62 from a competitor; eleven parties signed confidentiality agreements. The rival in the final round, "Party C," had demanded and received reimbursement of up to $7.5 million of expenses. MoffettNathanson put CyrusOne's enterprise value at $14.8 billion. CyrusOne's closing release is dated March 25, 2022. Advisers: Goldman Sachs, Barclays, Wells Fargo, Citigroup and J.P. Morgan for KKR and GIP; Morgan Stanley and DH Capital for CyrusOne.

There's a footnote to the CyrusOne story that I enjoy. As I wrote in Bandwidth, Brookfield Infrastructure won a bidding war with Macquarie in 2020 to buy Cincinnati Bell, the old phone company that had spun CyrusOne off. Both the parent and the child ended up owned by infrastructure funds.

American Tower and CoreSite

The CoreSite deal was the strangest of the four, and the one that ties most directly back to Bandwidth.

It didn't start as a bid. According to CoreSite's filing on the tender offer, on March 5, 2021, an American Tower executive, Steve Baker, emailed CoreSite's CEO, Paul Szurek, about "deploying data centers on American Tower real estate." It was a partnership idea for edge computing. In late July, the CEO of another communications-infrastructure company, Party A in the filing, raised a merger over breakfast, and in August it offered all stock, worth about $161.72 a share. In August, Szurek asked American Tower's Tom Bartlett whether his company wanted to do a deal. It did. In October, American Tower bid "in the area of $165.00." On the day the final bids came in, November 11, Del Deo wrote that CoreSite's "unique assets and a tower operator described as expressing interest" set it apart. That day American Tower bid $170 a share, all cash, against Party A's package worth about $164.58.

American Tower paid $170 a share, about $10.4 billion including debt, for CoreSite, the Denver REIT that had gone public at $16 in 2010. That's more than ten times the IPO price in eleven years. CoreSite had twenty-five data centers, twenty-one on-ramps to the cloud providers and more than 32,000 interconnections in eight U.S. markets. Its annualized revenue was $655 million and its adjusted EBITDA $343 million.

Let's do the arithmetic. $10.4 billion divided by $343 million of annualized EBITDA is about thirty times.The revenue and EBITDA figures are CoreSite's own, for the third quarter of 2021, annualized, as stated in the merger release. The merger release gave about $10.1 billion including debt; American Tower's later 10-Q put the total consideration at about $10.4 billion including assumed and repaid debt, and I have used the filing. On the release's $10.1 billion, the multiple is about twenty-nine times. MoffettNathanson's Del Deo and Michael Srour also put the enterprise value at about $10.4 billion and the multiple at about thirty times their estimate of the next twelve months' EBITDA, adjusted for construction in progress and backlog. Bloomberg put CoreSite's prior Friday close at $166.59 a share. The scene comes from CoreSite's Schedule 14D-9, November 29, 2021; it doesn't name Party A. After losing, Party A's lawyers asked for its information to be returned or destroyed.

Why does that number matter to me? In Bandwidth, I told how a talented young MoffettNathanson analyst named Nick Del Deo initiated coverage of Zayo during our 2014 IPO road show with a report comparing dark fiber to towers. He called towers "almost universally viewed as the best business model in telecom, if not one of the greatest business models of all time." Towers traded at about twenty-five times their profitability. Bandwidth companies like Zayo traded at eight to twelve times. Del Deo's verdict on whether dark fiber deserved the tower multiple was "close but no cigar."

Seven years later, a tower company paid a tower-plus multiple for a data-center company, and the same analyst was there to price it. On the morning of November 15, 2021, Del Deo opened his note, "This morning, something very unusual transpired: two acquisitions in one day." By his own adjusted measure of the next twelve months' EBITDA, American Tower was paying about thirty times for CoreSite, and KKR and GIP about twenty-three times for CyrusOne. He called the American Tower deal "far more interesting." Two years later, in a note titled "Data Centers through a Tower Lens," Del Deo was running data centers, not dark fiber, through the tower test.

And it was American Tower that did it. As I wrote in Bandwidth, of the three big U.S. tower companies, two, American Tower and SBA, "decided to stick to their knitting" while Crown Castle spent more than $11 billion buying fiber. Now American Tower had left its knitting too, just in a different direction. Tom Bartlett of American Tower explained it as a bet on the edge of the network: "We are in the early stages of a cloud-based, connected and globally distributed digital transformation." The idea was that wireless networks and data centers would converge, and the company that owned both the towers and the buildings where the networks met would win.

The deal had one more connection to this chapter's cast. Eight years earlier, in 2013, American Tower had paid $4.8 billion for Marc Ganzi's Global Tower Partners. The tower company that bought Ganzi's towers was now following Ganzi into data centers.

Was he right? Elliott Investment Management's later campaign against Crown Castle's fiber strategy, which I described in Bandwidth, is a warning about tower companies straying from towers. But data centers aren't fiber, and CoreSite's interconnection business looked a lot more like Equinix than like Crown's fiber. The jury was still out when this chapter closes. The deal closed on December 28, 2021. "We are thrilled to welcome CoreSite's exceptional team to American Tower," Bartlett said.

Let's put the wave in perspective. In February 2000, at the top of the Gold Rush, Exodus had a Market Value of more than $20 billion. Twenty-one months later, twenty-eight of its data centers sold for $560 million in cash. Twenty years after that, four data-center companies sold for about $46 billion in eleven months. The difference wasn't the buildings. Exodus's buildings were fine too. The difference was the tenants. Exodus rented to dot-coms that could vanish in a quarter. QTS, CyrusOne, CoreSite and Switch rented to the cloud companies and to enterprises that had moved their computing into somebody else's building for good.

The Infrastructure Funds—Stonepeak, EQT, Brookfield and Macquarie#

In Bandwidth, the infrastructure funds got a long footnote: a list of the fiber companies they bought between 2014 and 2020. In this book, they get a section, because by 2022 they owned a large share of the data-center industry.

The list was long, and many of the names will show up again here. In 2017, KKR bought 40 percent of Telxius, Telefonica's fiber spinout. In 2018, EQT funded a European roll-up of IP-Only, GlobalConnect and Broadnet. In 2019, Macquarie Infrastructure Partners bought Bluebird Network. In 2020, Stonepeak bought Level 3's South American business and branded it Cirion, and Digital Colony bought a majority of Beanfield. The funds learned digital infrastructure on fiber. Then they turned to the buildings at the end of the fiber.

Why did they pay so much? I explained the logic in Bandwidth using EQT as the example. The investors behind the infrastructure funds were mostly European pension and sovereign wealth funds, and at the time they were earning near-zero returns on their bonds. Infrastructure was seen as moderate risk. So they decided a mid-to-high single-digit return was a prudent trade-off. That lower cost of capital let them pay prices that were materially higher than anyone else would pay. If one bidder needs a high-single-digit return and the other needs a mid-teens return, the second bidder is going to lose the auction.

That is exactly what happened in data centers.

Stonepeak and Cologix

In February 2017, Stonepeak Infrastructure Partners bought a majority of Cologix and committed $500 million to grow it. Cologix had twenty-four data centers in nine U.S. and Canadian markets and about 1,600 customers. Its CEO and chairman was Grant van Rooyen, one of my colleagues from Level 3.Bandwidth spells his name "Van Rooyen"; Cologix's own releases and Data Center Frontier spell it "van Rooyen." The $3.0 billion recapitalization is from the title of Cologix's April 6, 2022 release as listed in Bandwidth's sources. Existing investors, including Columbia Capital, kept a material interest.

Stonepeak had been founded only in 2011 and managed more than $6.4 billion at the time of the Cologix deal. By the time it was bidding on Zayo a year and a half later, as I wrote in Bandwidth, it had $50 billion to invest. As I also wrote there, Cologix "had a great outcome for investors." In April 2022, Cologix announced that Stonepeak had completed a $3.0 billion equity recapitalization of the company.

EQT and EdgeConneX

EQT came next. In August 2020, five months after it closed on Zayo, EQT Infrastructure IV agreed to buy EdgeConneX from an investor group led by Providence Equity Partners. EdgeConneX operated or was developing more than forty facilities in thirty-three markets across North America, Europe and South America. The price wasn't disclosed.

The EQT partner quoted on the deal was Jan Vesely, the same Jan Vesely who had led EQT's side of the Zayo deal. "EQT has followed EdgeConneX's journey from its early years to its growth into a top data center industry player," he said. The fund that bought Zayo's fiber was now buying the buildings that fiber connected.

The Telcos Sell

The funds' biggest source of buildings was the phone companies.

In Chapter V, the telcos bought data centers at the top of the Phoenix. Verizon bought Terremark in 2011 for about $1.4 billion of equity. CenturyLink bought Savvis the same year for about $2.5 billion of equity, or about $3.2 billion including debt. The pitch both times was that a phone company with data centers and a cloud could sell more to its enterprise customers.

Five years later, both sold, and AT&T sold alongside them.

We've seen Verizon's sale to Equinix. In November 2016, CenturyLink agreed to sell fifty-seven data centers to a consortium of BC Partners, Medina Capital Advisors and Longview Asset Management for $2.15 billion in cash plus a minority stake in the new company valued at $150 million. The new company had more than 3,500 customers and 2.6 million square feet of raised floor.The new company was named Cyxtera Technologies. The deal closed on May 1, 2017, and the fifty-seven data centers had about 195 megawatts of power. CenturyLink's fifty-seven data centers and Savvis's portfolio aren't the same set of buildings, so the 2011-versus-2016 comparison in the text is directional, not exact. Its CEO would be Manuel Medina, the founder of Terremark, who had sold his own company to Verizon five years earlier.

Here's the part that hits home. CenturyLink used the proceeds partly to fund its $24 billion acquisition of Level 3, the company I co-founded. The data centers CenturyLink had assembled, Savvis's among them, helped pay for the fiber network I had spent the Boom building.

Then AT&T sold. In a deal announced in mid-2018 and closed at the start of 2019, Brookfield Infrastructure and its partners paid $1.1 billion for thirty-one AT&T data centers, eighteen in the United States and thirteen elsewhere, with more than a thousand customers. They named the company Evoque Data Center Solutions. AT&T said it would use the proceeds to bring its debt down to about 2.5 times EBITDA by the end of 2019.The strategic alliance was announced in 2018. AT&T announced on January 1, 2019, that the sale had closed. Evoque joined AT&T's colocation ecosystem program, so AT&T could keep reselling the space.

Three phone companies, three exits, all for the same reason: they needed the money for their core businesses, or to pay down debt. As I wrote in Bandwidth, the big phone companies are "stuck in deep mud" when it comes to separating their bandwidth businesses from their core. Data centers were easier to separate. So they went first.

Macquarie

Macquarie took two routes into data centers. In April 2018, Macquarie Infrastructure Partners invested in Aligned Energy, a developer whose first data center opened in Plano, Texas, in 2015, sharing control with BlueMountain Capital. The CEO, Andrew Schaap, had come from Digital Realty. Aligned's cooling technology let customers like eBay, Lenovo and Telus scale up and pay only for the capacity they used, an idea that should sound familiar from the AWS section of Chapter VI.

Then, in April 2020, a Macquarie-led consortium bought 88 percent of AirTrunk, a hyperscale developer in the Asia-Pacific region, at a valuation above A$3 billion. AirTrunk had five data centers and more than 450 megawatts when fully developed. Its founder, Robin Khuda, kept a material stake and stayed on as CEO. "AirTrunk has established itself as the dominant player in the APAC hyperscale market," he said. Macquarie's Frank Kwok gave the investor's view: "Our investors are attracted to hyperscale data centre providers such as AirTrunk."Macquarie's September 2024 release on AirTrunk's sale names PSP Investments as its co-owner.

The Same Names

Look back at the firms that competed to buy Zayo in 2018 and 2019: Blackstone, I Squared, KKR and Stonepeak, with Charlesbank and GTCR, on one side, Digital Colony and EQT on the other. Now look at who owned the data-center industry by 2022. Blackstone owned QTS. KKR co-owned CyrusOne. Stonepeak owned Cologix. EQT owned EdgeConneX. DigitalBridge owned DataBank, Vantage and, by the end of the year, Switch. Brookfield, which lost Switch to DigitalBridge, owned Evoque.

The cast of the Zayo auction had become the owners of the data-center industry. It was a small club, and its members had all done the same math.

On the Eve#

By November 2022, the data-center business had been through a full Reconsolidation.

Two public giants stood on top. Digital Realty had absorbed Telx, DuPont Fabros and Interxion. Equinix had absorbed Telecity and Verizon's data centers and had found in xScale a way to serve the cloud companies with a sovereign fund's money. Of the wholesale REITs that went public in Chapter VI, none was left on the stock market. Private owners held most of the rest: Blackstone, KKR and GIP, American Tower, DigitalBridge, Stonepeak, EQT, Brookfield and Macquarie. The phone companies were out.

The new owners had paid prices that would have been unthinkable in 2002. They paid them for one reason: cash flow. A data center leased to a cloud giant for ten or fifteen years produced steady, rising rent from the best credit in the world. To an infrastructure fund, or to a perpetual-capital vehicle, that stream of cash looked as close to Buffett's "between now and doomsday" as anything in real estate. The 2008 crisis had proved that the tenants kept paying. The 2009 rebound had proved that investors would notice. And for investors earning near-zero returns on their bonds, a mid-to-high single-digit return on a data center looked like a bargain.

Did the high prices pay off for the companies that did the building? Years later, MoffettNathanson went back and measured. Its last industry-wide study of returns on invested capital used results through 2021, and it was the last time the study could be done across many operators, because CoreSite, CyrusOne, QTS and Switch had all been taken off the market. The answer was sobering. Operators without an interconnection business earned about their cost of capital, and no more. Equinix and CoreSite did better. "Absent a gaping valuation disparity," Del Deo wrote, "interconnection was the place to be." Digital Realty's return on capital fell more than 3 percentage points from 2018 to 2023, as prices and development yields fell.

What could go wrong? The bears had two answers. The first was concentration: a handful of tenants, each of whom could build its own buildings, and each of whom knew exactly what a building cost. The second was valuation: CoreSite at thirty times EBITDA had been priced like a tower, and towers don't need new buildings every few years to keep growing. The bulls had one answer, and it was the one Ganzi had given in 2017. It was still the early innings.

In the summer of 2022, both sides got a headliner. In July, the short seller Jim Chanos told the Financial Times he was raising money to bet against data-center REITs, naming Digital Realty and Equinix. The cloud, he said, was "their enemy, not their business," and "Their three biggest customers are becoming their biggest competitors." Bill Stein answered on CNBC: "I think Jim maybe isn't aware that demand has never been stronger in our space."

The same month, the thing that would decide the argument showed up, and it wasn't a tenant. In July 2022, Dominion Energy told its large customers that "transmission constraints in a pocket of eastern Loudoun County" would delay new connections. Some new data centers there might not get connected until 2026. Andy Power of Digital Realty called it "a bottleneck in a portion of Ashburn," not "a permanent feature." Stein predicted that "prices are going to go up, because supply is constrained." Ganzi called it "a bit of a wake-up call for the region."The Chanos and Stein quotes are from Bisnow, July 7, 2022, which also quotes Wells Fargo's estimate that the cloud companies leased about 60 percent of their needs in 2022. Dominion's notice and Power's quotes are from Data Center Frontier, July 29, 2022; Stein's and Ganzi's from Data Center Frontier, September 26, 2022, which reported that the constraints touched two-thirds of Loudoun's substations, with relief expected from a new 500-kilovolt line in 2026. The returns study is Del Deo and Chris Chae, "Data Centers: Do ROIC Trends Match the Narrative of the AI Boom?" MoffettNathanson, April 13, 2026. MoffettNathanson uses its own lifecycle depreciation rather than reported depreciation.

Neither side was forecasting what came next.

On November 30, 2022, OpenAI released ChatGPT.

Days later, in early December, DigitalBridge and IFM closed on Switch for about $11 billion.The closing release, "DigitalBridge and IFM Investors Complete $11 Billion Take-Private of Switch," is dated December 6, 2022, the day the deal closed. It was the last big deal of the Reconsolidation, and its buyers had priced it the way everyone in this chapter had priced data centers: as a steady stream of cash from a mature industry.

Was a data center a bond, or was it about to become the most contested real estate in technology? And if Ashburn was a preview, what would the buildings run short of first, space or power?

In Bandwidth, I borrowed Buffett's warning about the party where the clocks have no hands. The music was about to start again.

Table 7.1: The Reconsolidation's Data-Center Deals, 2015–2022#

In June 2024, Elon Musk's artificial-intelligence company, xAI, chose an empty building in South Memphis, Tennessee. It had been an Electrolux factory. In early September, Musk posted that a cluster of 100,000 Nvidia H100 chips was running inside it. "From start to finish, it was done in 122 days," he wrote.The 122-day figure is xAI's own claim, from Musk's post in early September 2024; it has not been independently verified. xAI chose the Memphis site in June 2024. According to R&D World, xAI doubled it to 200,000 chips in ninety-two more days. Musk called it Colossus.

The local utility couldn't deliver power that fast, so xAI brought its own. Gas turbines arrived on trailers and lined up outside the plant. By the spring of 2025, according to the Southern Environmental Law Center, at least thirty-five of them had been running for about a year without air permits, with up to 421 megawatts of capacity. The county health department had told reporters the turbines were exempt. A satellite image from June 15, 2025 showed at least twenty-six of them still on site. By the law center's count, the turbines could emit more than 2,000 tons of nitrogen oxides a year, against 1,077 tons for the county's biggest listed source, Memphis International Airport.

The neighbors didn't need a satellite. Boxtown, the nearest neighborhood, is nearly all Black. Freed slaves settled it, and it got its name from the houses they built out of boxcar scraps. It sits near the Tennessee Valley Authority's retired Allen coal plant and its coal ash. "They always come to southwest Memphis," KeShaun Pearson of Memphis Community Against Pollution told NBC News. Memphis's mayor, Paul Young, saw it differently. He called xAI "a tremendous opportunity."

Here is what strikes me about that scene. In Chapter III, the founders of Exodus knocked out a ceiling and rolled in portable air conditioners to keep forty servers alive in a Sunnyvale office suite. Three decades later, the most aggressive computing company in the world was rolling in portable power plants. The problem was the same one: keeping the machines powered and cool. Only the zeros had changed.

Chapter VII ended on November 30, 2022, the day OpenAI released ChatGPT. Within two months, UBS estimated, it had 100 million monthly users, the fastest ramp of any consumer app on record. What followed was the second Gold Rush in this book, and the biggest. Nvidia sold the picks and shovels and briefly became the most valuable company on earth. OpenAI and Microsoft, and then Masayoshi Son and Larry Ellison, promised half a trillion dollars of data centers from a podium at the White House. CoreWeave, a crypto miner from New Jersey, became this cycle's Exodus and maybe its Global Crossing. Amazon, Alphabet, Meta and Microsoft spent more on buildings and chips in a single quarter than the whole fiber industry spent in a year at the top of the last Boom. Three Mile Island was brought back to life. Voters pushed back. And in the last week of October 2026, Anthropic filed to go public.

The Reconsolidation was over. The music had started again.

Nvidia—Jensen Huang#

Every gold rush makes its surest fortune for whoever sells the shovels. In this one, the shovel was a graphics chip, and the shopkeeper was Jensen Huang.

Huang co-founded Nvidia in 1993 to make graphics chips for video games. Drawing a video game means doing thousands of small calculations at the same time. So does training a neural network, as AI researchers discovered in the early 2010s, and Nvidia's software let them program its chips for the job.

Nvidia's own filings show how early the bet started and how long it took to pay. Nvidia began reporting a "Datacenter" line in its fiscal 2016 annual report, recast back two years. In fiscal 2014 that line was $199 million. The next year's report credited "strong demand for deep learning training for AI" and named Facebook, Amazon and Baidu as customers. But as late as the year ended January 2020, Data Center revenue was $2.98 billion, "up 2% from a year ago," held back by "lower hyperscale sales." As recently as fiscal 2016, three-quarters of Nvidia's revenue still came from products tied to the PC.

The mechanics of Nvidia's edge are simple to state and hard to copy. An AI data center is a building full of Nvidia chips, wired together so tightly they act like one machine. Nvidia sells the chips, the boards, the networking between them, and increasingly whole racks. It reports all of that as "Data Center" revenue.

Here is the one number. In the fiscal year that ended in January 2023, two months after ChatGPT launched, Nvidia's Data Center revenue was $15.0 billion.Nvidia's fiscal 2023 Form 10-K states Data Center revenue of $15.01 billion. Nvidia's fiscal years end in late January. The full series from its 10-Ks, in millions of dollars, runs 199, 317, 339, 830, 1,932, 2,932, 2,983, 6,696, 10,613, 15,005, 47,525, 115,186 and 193,700 for fiscal 2014 through fiscal 2026. Fiscal 2021's jump reflects the Mellanox acquisition, completed in April 2020 for $7.13 billion, which contributed 10 percent of that year's total revenue. The next year it was $47.5 billion. Then $115.2 billion. Then, in the year ended January 2026, $193.7 billion. From $199 million to $193.7 billion is roughly a thousandfold in twelve years. In the quarter that ended in July 2026, Nvidia booked $89.0 billion of Data Center revenue in three months, six times what it had booked in all of fiscal 2023, at a 75.0 percent gross margin. In February 2026, Nvidia reported quarterly net income of $43.0 billion. For the year, it returned $41.1 billion to its shareholders.

When the first wave hit, Huang put it plainly: "Accelerated computing and generative AI have hit the tipping point." By the time of his Computex keynote in Taipei in May 2025, he had a name for what his customers were building. They were "AI factories," and their product was tokens, the small chunks of text an AI model reads and writes. "AI is now infrastructure," Huang said, "and this infrastructure, just like the internet, just like electricity, needs factories." Then he added, about his customers' buildings, "They're not data centers of the past."

I'd put it differently. They were the glass house again. In the 1960s, the computer was a giant machine in a sealed room, and companies rented time on it. In this book, that glass house came back as hosting, then as the cloud. Now it came back as a factory, one giant computer to a building, and you rented its output by the token.

The Trillion-Dollar Club

Wall Street did the arithmetic fast. On May 30, 2023, Nvidia became the first chipmaker to touch a $1 trillion Market Value, during the trading day.The $1 trillion mark on May 30, 2023 was touched during the trading day; Business Insider and Bloomberg Law report that Nvidia closed just below it. The first close above $1 trillion, and the dates of the $2 trillion milestone (widely reported as February 23, 2024 intraday and March 1, 2024 at the close), are not verified. On July 9, 2025, AP reported a close of $3.972 trillion, just short of $4 trillion. On June 5, 2024, it closed at $3.019 trillion and passed Apple. On July 9, 2025, it became the first company in history to touch $4 trillion. On October 29, 2025, about four months later, it became the first to reach $5 trillion, closing at $5.03 trillion.

That's a long way from video games.

The bulls had the numbers on their side. In November 2025, Nvidia's chief financial officer, Colette Kress, said the company had "visibility" to $500 billion of revenue from its Blackwell and Rubin chips through calendar 2026. By August 2026, Huang was making a claim that sounded like a business school lecture: "AI has reached its inflection point. It's doing useful work. Its tokens are productive and profitable. Now, compute is revenue."

The skeptics heard a different sentence in the same call. "This time last year," Huang said, "one lab alone was driving the buildout." He meant it as progress: the customer base had broadened. But it was also an admission of how concentrated the demand had been. And some of those customers were Nvidia's own investments. As we'll see, Nvidia anchored CoreWeave's IPO, put $30 billion into OpenAI's 2026 funding round, and agreed to invest up to $10 billion in Anthropic. The shovel seller was also financing the miners.

Election Week

On November 3, 2026, Election Day, Nvidia closed at $206.69 a share, a Market Value of $4.99 trillion. It was the most valuable company in the world, ahead of Apple at $4.84 trillion. The next day Apple took the title back. On November 5, Nvidia fell 9 percent and lost about $374 billion of Market Value in a single session. As I write, on November 6, Nvidia ranks second, at about $4.5 trillion.The ranking changed hands three times in a week. Apple took the lead at the October 30 close, after Nvidia fell 4.2 percent, and held it for two sessions. On November 3, Nvidia closed at $4.99 trillion against Apple's $4.84 trillion, Microsoft's $3.87 trillion and Alphabet's $3.44 trillion (24/7 Wall St.). On November 4, Apple hit a record $354.50 a share and $5.08 trillion, against Nvidia's $4.83 trillion. CompaniesMarketCap on November 6 showed Apple at $4.975 trillion and Nvidia at $4.506 trillion, from delayed quotes with no timestamp. Re-check at publication.

Nvidia's next earnings report is expected around November 17. Its profits are the one thing in this chapter nobody disputes. The question is whose budgets they come from, and for how long.

OpenAI, Microsoft and Stargate—Sam Altman, Masayoshi Son and Larry Ellison#

"In 20 years following the internet space, we cannot recall a faster ramp in a consumer internet app," UBS's analysts wrote in February 2023. ChatGPT had reached about 100 million monthly users in January, two months after launch. TikTok had taken about nine months. Instagram had taken about two and a half years. OpenAI started charging $20 a month for a premium version.

Sam Altman ran OpenAI, which had started as a nonprofit research lab in 2015. His company's real constraint was never users. It was computing power, and computing power meant data centers.

Microsoft was first to the table. In January 2023, it announced the third phase of its partnership with OpenAI, a "multiyear, multibillion dollar investment." It didn't disclose the amount; the press reported $10 billion.Microsoft held 32.5 percent of OpenAI before the recapitalization and "roughly 27 percent on an as-converted diluted basis" after it. Under the October 2025 terms, Microsoft's rights to OpenAI's technology run through 2032, and any declaration that OpenAI has reached artificial general intelligence must be verified by an independent panel. The key sentence was about real estate: "As OpenAI's exclusive cloud provider, Azure will power all OpenAI workloads across research, products and API services." Microsoft had made itself OpenAI's landlord.

Stargate

Two years later, the landlord had company. On January 21, 2025, OpenAI, Oracle, SoftBank and an investor called MGX stood at the White House and announced Stargate: up to $500 billion over four years for AI infrastructure in the United States, starting with $100 billion. SoftBank would handle the finances, and Masayoshi Son would be chairman. OpenAI would run the operations.

Son set the tone. "This is the beginning of our golden age," he said. Larry Ellison, Oracle's co-founder, wanted everyone to know the money was more than a promise: "The data centers are actually under construction. The first of them are under construction in Texas."

Readers of Bandwidth have met two of these partners before. In September 1999, Global Crossing announced a $1.3 billion joint venture with Microsoft and SoftBank to build a fiber network across Asia. Microsoft and SoftBank each committed to buy $200 million of Global Crossing's network capacity. Brad Feld, who co-founded SoftBank's venture fund in 1996, wrote in his foreword to Bandwidth that he rode that boom to its 2000 peak "and then suffered mightily." Son himself lost most of his fortune in the Dot Bomb. Twenty-five years later, he was back on the dance floor, and he was calling it a golden age.

Within hours, Elon Musk, who had co-founded OpenAI and then fallen out with Altman, posted that Stargate's backers "don't actually have the money" and that SoftBank had "well under $10B secured."Musk's words are taken from syndicated CNN and CNBC coverage, not from his original post; Fortune and Data Center Knowledge quote the same words. I haven't seen the post itself. Altman replied that Musk was wrong and invited him to come see the first site.

Abilene

The first site was in Abilene, Texas. Construction had started in July 2024, before Stargate had a name. The arrangement was layered like a wedding cake. A company called Lancium owned the land. Crusoe, the developer, built the buildings and leased them to Oracle. Oracle filled them with Nvidia chips and rented the computing to OpenAI. Readers of Chapter VI will recognize the wholesale data-center model, with one more layer of tenant on top.

The plan was eight buildings and about 1.2 gigawatts. A gigawatt is a thousand megawatts, a bit more than the output of the Three Mile Island reactor we'll meet in the Power section. In May 2025, Crusoe raised $11.6 billion to go from two buildings to eight. By Oracle's September 2026 earnings call, according to tracker coverage, six of the eight buildings, 618 megawatts, had been delivered. That is fast. It is also behind the speed of the press releases.

In September 2025, Stargate named five more sites, in Texas, New Mexico, Ohio and the Midwest, bringing the plan to about 7 gigawatts and more than $400 billion over three years. "AI can only fulfill its promise if we build the compute to power it," Altman said. Six months later, in March 2026, Bloomberg reported that Oracle and OpenAI had dropped plans to expand the Abilene campus, citing financing complications and OpenAI's shifting demand forecasts.The DatacenterDynamics story was blocked to automated reading; its content comes from search excerpts citing Bloomberg. Trackers disagree on whether the dropped expansion was about 600 or about 900 megawatts. One account says OpenAI did not want to mix Blackwell and Vera Rubin chips at one site. Meta was reported to be in talks, with Nvidia's help, to pick up the capacity.

Shifting demand forecasts. Hold that phrase until the Bubble section.

Ellison's Morning

The biggest winner of 2025, on paper, was Larry Ellison. On September 9, 2025, Oracle reported $455 billion of remaining performance obligations, the contracted revenue it had not yet delivered. That was up 359 percent from a year earlier, and it came from four multibillion-dollar contracts with three customers. The next day, Oracle's stock rose about 36 percent, its best day since 1992. For a few hours, Bloomberg ranked Ellison as the richest person in the world, at about $393 billion, ahead of Musk at about $385 billion.

By August 2026, Oracle's backlog had grown to $664.0 billion. So had its debt. Oracle's filings show debt of $92.9 billion in May 2025 and $130.1 billion a year later.The debt figures are the carrying amounts as tagged in Oracle's 10-K data; the scope of the tag should be checked before print. Reuters coverage reported that Oracle planned about $50 billion of stock and bond sales for 2026 and that Morgan Stanley saw $100 billion or more of further needs through the first half of 2028. Oracle's five-year credit default swaps reportedly reached about 151 basis points in 2026. Its OpenAI contract, worth about $300 billion, was described in Reuters coverage as "unprecedented single customer revenue exposure." The price of insuring Oracle's bonds against default, measured by its five-year credit default swaps, climbed to the highest level since 2009.

A backlog is a promise to pay. In Bandwidth, I wrote about companies that booked the promises and borrowed against them. Oracle is a far stronger company than any of those were, with a real software business underneath. But the shape is familiar: a builder taking on debt to serve one customer whose own money comes from investors.

The Landlord Moves Out

By late 2025, OpenAI had outgrown its first landlord. In October 2025, after OpenAI's recapitalization, Microsoft announced that it held "roughly 27 percent" of OpenAI, worth about $135 billion. OpenAI "has contracted to purchase an incremental $250B of Azure services." In return, Microsoft gave up its right of first refusal to be OpenAI's computing provider. The exclusive was over.

In March 2026, OpenAI raised $122 billion at an $852 billion valuation. SoftBank put in $30 billion. So did Nvidia. Amazon put in $50 billion, $35 billion of it contingent on an IPO or on OpenAI declaring artificial general intelligence. Every large check came from a company that also sold OpenAI chips, cloud capacity or both.

The same month, the Department of Energy announced a plan for the former Portsmouth Gaseous Diffusion Plant in Piketon, Ohio, a Cold War uranium site. It would host a 10-gigawatt data center and up to 10 gigawatts of new generation, 9.2 gigawatts of it natural gas. SoftBank's SB Energy would partner with the local utility, AEP Ohio, and the Energy Department cited $33.3 billion of Japanese funding for the gas plants under a U.S.–Japan trade agreement. Energy Secretary Chris Wright said the project would "ensure the United States wins the AI race." In August 2026, according to CNBC, OpenAI announced that it had secured 8 gigawatts at the Ohio campus, which CNBC called the world's largest planned AI campus, at a possible cost above $500 billion. One site now carried the price tag of the original Stargate.

Microsoft's own report card came in on October 28, 2026. Azure revenue, disclosed in dollars for the first time, grew 48 percent to $33.17 billion in the quarter. Microsoft also recorded a loss of about $200 million on its OpenAI investment. Its chief financial officer, Amy Hood, gave the bull case in six words: "We still have demand that exceeds supply."

Musk's question in January 2025 had been rude, but it was the right one. Who has the money? By 2026, the answer was clear. Mostly, the suppliers did.

CoreWeave—Michael Intrator#

Every Gold Rush has a company that captures the whole thing in one stock chart. In the hosting boom, it was Exodus. In the fiber boom, it was Global Crossing. In the AI boom, it is CoreWeave.

CoreWeave began as something else entirely. According to its IPO prospectus, it was "formed in September 2017 as a Delaware limited liability company under the name The Atlantic Crypto Corporation LLC." The founders, led by Michael Intrator, used graphics chips to mine cryptocurrency, a business the prospectus says the company later discontinued. In December 2019, the company renamed itself CoreWeave and pointed the same kind of chips at a different customer: anyone who needed to rent GPU computing by the hour. It kept its headquarters in Livingston, New Jersey.

The edge was simple. When ChatGPT hit, Nvidia's newest chips were the scarcest commodity in technology, and CoreWeave had a place near the front of Nvidia's line. The big clouds were building their own capacity as fast as they could, and it still wasn't enough. So they rented CoreWeave's.

The one number is revenue. CoreWeave booked $15.8 million in 2022. In 2023 it booked $228.9 million. In 2024 it booked $1.915 billion, a 120-fold increase in two years. By the end of 2024, it ran thirty-two data centers with more than 250,000 GPUs on more than 360 megawatts of active power, and had about 1.3 gigawatts under contract. Its business was filling data centers with chips and renting out the chips; the buildings themselves were leased from others.

Look closely at who was paying. Microsoft was 35 percent of CoreWeave's revenue in 2023 and 62 percent in 2024, and the top two customers together were about 77 percent. CoreWeave also signed a contract with OpenAI worth up to about $11.9 billion through October 2030. And the prospectus showed how CoreWeave paid for its chips: with loans carrying effective interest rates of 10 to 17 percent. One delayed-draw term loan cost 15 percent. CoreWeave's debt principal went from $1.56 billion at the end of 2023 to $8.03 billion a year later.

Nvidia sat on every side of the table. It owned more than 5 percent of CoreWeave. It supplied, in the prospectus's words, "all of the GPUs used in our infrastructure today." And it was a customer: under a 2023 agreement, "NVIDIA has paid us an aggregate of approximately $320 million." The founders, meanwhile, had already taken money off the table. In two private tender offers before the IPO, Intrator and two of his co-founders sold shares for a total of about $488 million.CoreWeave, Inc., prospectus, March 31, 2025. The top three customers were about 41 percent of revenue in 2022 and 73 percent in 2023. The December 2023 tender, at $15.50 a share, paid entities affiliated with Intrator $110.0 million; the 2024 tender, at $47.00, paid them $50.0 million. Co-founders Venturo and McBee received about $176.9 million and $150.9 million across the two. The prospectus also described CoreWeave's $7.6 billion delayed-draw facility, led by Blackstone and Magnetar, as "one of the largest private debt financings in history."

The IPO

CoreWeave priced its IPO on March 27, 2025, at $40 a share. The bankers had marketed it at $47 to $55. They also cut the number of shares from 49 million to 37.5 million. The deal raised $1.5 billion, which CNBC still called "the biggest U.S. tech offering since 2021." At $40, CoreWeave was valued at about $19 billion, against about $26.5 billion at the top of the range. Nvidia reportedly anchored the deal with a $250 million order.The $19 billion figure is on a basic share count; the fully diluted value was higher. Nvidia's $250 million order is reported, not confirmed by CoreWeave. The first-day trading figures are from CNBC.

Compare that with Exodus in Chapter III. Exodus priced at $15, opened at $24.75 and closed its first day up 84 percent. CoreWeave opened at $39 and closed its first day flat. The 1998 crowd was buying a forecast. The 2025 crowd was buying $1.9 billion of real revenue, an $863 million net loss and one very large customer, and it bought cautiously.

Then the stock took off. On June 20, 2025, it touched $187 during the trading day and closed at a record $183.58.

The analysts split. Gil Luria of D.A. Davidson cut CoreWeave to Underperform in May 2025: "Investors regretted scaling WeWork, and they may not want to scale this business." MoffettNathanson's Del Deo started coverage at Neutral the same month and stayed there. His worry was the buildings. CoreWeave's "staggering data center lease commitments" were more than five times what it had in service at the end of 2024. In June 2025, in a note titled "No Country for Old Megawatts," he added it up: 420 megawatts of active power and leases committed for another 2,100. "Chips are useless without data centers in which to house them," he wrote. By his count, a company that had taken its current name only in 2019 had more power in hand or under contract than Digital Realty's whole global portfolio.

The Numbers Get Big

By the summer of 2026, CoreWeave had become what it set out to be. In the second quarter, revenue more than doubled, up 112 percent to $2.58 billion. Its backlog of signed contracts was $104 billion, before more than $25 billion of new third-quarter commitments. Meta signed for an additional $21 billion. Anthropic signed up. So did the trading firm Jane Street, for $6 billion. CoreWeave ran 1.5 gigawatts of active power and guided to $12.4 to $13.2 billion of revenue for 2026. Intrator had reason to brag: "Pricing and margins for our Blackwell and Vera Rubin SKUs are setting new highs."

To get there, CoreWeave guided to $35 to $39 billion of capital spending in 2026, about three times its revenue. It lost $626 million in the second quarter. It raised about $18 billion in that quarter alone. CNBC put its debt at $35 billion; an analysis by Trefis put total debt at $51.6 billion at the end of June, about 193 percent of the company's equity Market Value.CNBC put CoreWeave's debt at $35 billion; Trefis put total debt at $51.61 billion at June 30, 2026. The definitions differ, and the 10-Q should settle it before print. Benzinga cited a backlog of about $129 billion including third-quarter commitments. Trailing free cash flow was negative $13.66 billion. By October 2026, CoreWeave's bonds maturing in 2031 and 2032 reportedly yielded above 15 percent.

CoreWeave also tried to buy its way into more power. It made a roughly $9 billion all-stock offer for Core Scientific, another crypto miner turned AI host. In October 2025, Core Scientific's shareholders voted it down. It was CoreWeave's second failed approach, after one in June 2024.

From May 8 to November 4, 2026, CoreWeave's stock fell 57.6 percent, including a 9.9 percent drop the day after the midterms. It reports third-quarter results on November 10, four days after I write this.

Exodus, Global Crossing, or Neither?

The comparison is tempting, and I've been making it since the first line of this section. So let me be careful with it.

Here is the case that CoreWeave is this cycle's Exodus. Both companies rented space and services inside data centers to the hottest companies of their day. Both grew revenue at a pace no one had seen. Both relied on a handful of giant customers. Exodus had Yahoo and eBay. CoreWeave had Microsoft, then OpenAI and Meta. And both financed growth with debt. Exodus filed for bankruptcy owing about $2.9 billion of borrowed money in eight issues of notes.EXDS (Exodus) second amended disclosure statement, filed May 3, 2002, which counts the borrowed money in five series of senior notes and three convertible subordinated issues. Some accounts give about $3.5 billion in six bond issues; I have used the disclosure statement. At the end of 1999, Exodus had $1.53 billion of notes against $17.6 million of stockholders' equity.

Here is the case for Global Crossing. Gary Winnick raised $20 billion of debt and equity, as I wrote in Bandwidth, and built an asset whose value depended on demand that hadn't shown up yet. His suppliers and customers were often the same companies. Microsoft and SoftBank bought capacity from Global Crossing while investing alongside it. CoreWeave's supplier, Nvidia, was also its investor and the anchor of its IPO. Its biggest customer, Microsoft, ran a cloud of its own.

Now the other side, and it deserves equal time. Exodus had $12.4 million of revenue when it went public. CoreWeave had $1.9 billion. Global Crossing sold capacity on a glutted market at falling prices. CoreWeave, in mid-2026, was reporting record prices and margins on its newest chips. The bulls had a point about older chips too. According to Bernstein's Stacy Rasgon, five-year-old Nvidia A100s still earned "comfortable" margins, and CoreWeave re-leased three-year-old H100s within 5 percent of their prior price. And CoreWeave's earliest big contracts were signed with Microsoft and Meta, two of the most creditworthy companies on earth.

The newer ones were a different story. By late 2025, Del Deo noted, CoreWeave had diversified its backlog, an IPO-era worry, but largely through customers without investment-grade credit ratings, such as OpenAI and Poolside. By October 2026, he wrote, the other half of its demand was dominated by OpenAI, Anthropic and Jane Street.

The difference that worries me is the one that sounds most technical. In Chapter IV, I wrote that when Exodus went bankrupt, "the buildings were fine. The buildings were always fine." Twenty-eight of Exodus's data centers sold to Cable & Wireless for less than 3 percent of Exodus's peak Market Value, and the servers kept humming. CoreWeave's most valuable assets are chips, not buildings. A data center lasts for decades. A GPU lasts as long as nobody builds a better one. If the chips hold their value, as Rasgon's evidence suggests, CoreWeave's debt is backed by productive assets with paying tenants, some of them blue-chip. If they don't, the lenders will find out what a three-year-old GPU is worth in a downturn. In August 2026, Del Deo described the market for computing as "in backwardation (forward prices are lower than current prices)." In plain English, buyers expected compute to get cheaper. About 98 percent of CoreWeave's revenue was locked into multi-year take-or-pay contracts, so it couldn't be repriced down. It couldn't be repriced up, either, until the contracts ran out.

Is a GPU more like lit capacity, which got repriced to nothing after 2001, or like conduit, which held its value? Del Deo took on that question in an October 2025 note comparing the Telecom Bubble with the AI Boom. In telecom, most of the money went into fiber and conduit with decades of life, and the electronics were a minority. In AI, most of the money goes into GPUs that last years, and the buildings are the minority. Telecom capacity was built "on a speculative basis"; AI capacity is mostly built after a customer signs. "The fiber and conduit infrastructure deployed during the Telecom Bubble have long economic lives," he wrote. "In fact, much of it is supporting today's AI Boom!"

Which is CoreWeave? I don't know, and neither does the market, which is why the bonds yield more than 15 percent.

xAI Colossus—Elon Musk#

When we left Memphis at the start of this chapter, xAI had built a supercomputer in 122 days and powered it with gas turbines nobody had permitted. The permits came next, and then the lawyers.

On July 2, 2025, the Shelby County Health Department approved a permit for fifteen permanent turbines. The Southern Environmental Law Center appealed for the NAACP and a local group, Young, Gifted & Green. "The health department failed to hold xAI accountable for unlawfully installing 35 polluting methane gas turbines," said Amanda Garcia of the law center. On December 15, 2025, after a seven-hour meeting, the county's air-pollution board dismissed the appeal as moot.

Meanwhile, xAI was building Colossus 2 at a second Memphis site, a building of about 1 million square feet in Whitehaven. The Greater Memphis Chamber had told NBC News that xAI would not use gas turbines there. It didn't, in Tennessee. The turbines went south, across the state line, to Southaven, Mississippi.

In April 2026, the NAACP, represented by the Southern Environmental Law Center and Earthjustice, sued xAI and its subsidiary, MZX Tech, in federal court in northern Mississippi. The suit alleged twenty-seven unpermitted gas turbines in Southaven, feeding Colossus 2 across the border, with potential emissions of more than 1,700 tons a year of nitrogen oxides. The plaintiffs also objected to a proposed permanent plant of forty-one turbines on the same site. "A data center should not be a potential death sentence for a community's health," said Abre' Conner of the NAACP. "xAI isn't above the law," said Laura Thoms of Earthjustice.

Washington, meanwhile, moved in xAI's direction. In January 2026, the EPA created a new category for temporary turbines, allowing them to stay at a site for up to twenty-four months. In June 2026, the Justice Department moved to intervene in the Mississippi case and have it dismissed, citing national security.

On October 30, 2026, Judge Debra Brown ruled that the case could go forward. A turbine on a trailer, she found, is still a stationary source of pollution and not exempt mobile equipment. The plaintiffs now allege at least 495 megawatts of generation at what they call the "Colossus Gas Plant." Neil Osnato, founder of Persistence Analytics Group, quoted by TechTarget, summed up the ruling better than I can: "A trailer can make a turbine portable. It does not necessarily make its regulatory obligations temporary."

By then, xAI was no longer independent. In February 2026, Musk's rocket company, SpaceX, acquired it, and Bloomberg valued the combined company at $1.25 trillion. Musk's memo framed the deal around putting data centers in orbit. "Global electricity demand for AI simply cannot be met with terrestrial solutions," he wrote.

The final twist is in another company's IPO filing. Anthropic's registration statement lists "SpaceX.AI LLC (formerly known as X.AI LLC)," now a wholly owned SpaceX subsidiary, as one of its computing suppliers, for up to $120.5 billion. The man who said Stargate's backers didn't have the money had become a landlord to one of OpenAI's chief rivals.

Two things are true about Colossus. xAI built faster than anyone thought possible, and the speed was real. So were the turbines, and so were the neighbors who breathed what came out of them. The question for the industry is whether it can keep the first without the second.

The Capex Arms Race—Amazon, Alphabet, Meta and Microsoft#

Kevin O'Hara, my colleague at Level 3, is quoted in a 2025 IEEE Communications Society blog post summing up what the fiber Boom looked like from the inside: "We were spending about $110 million a week."

In the third quarter of 2026, Amazon spent $77.3 billion on property and equipment. That is about $6 billion a week, fifty-four times Level 3's pace at the top of the fiber Boom. It was one company, in one quarter.

The four big clouds spent capital in this cycle the way the fiber industry did in the last one, only more so. By one tally, the telecom carriers spent "over $100 billion" on fiber from 1998 to 2000. Here is what the cloud companies spent in cash on property and equipment, not counting leases:Figures are cash purchases of property and equipment from the companies' SEC filings, excluding finance leases. Amazon's line is "purchases of property and equipment." Microsoft's headline "capex including finance leases" is higher, and Microsoft changed its useful-life and lease accounting at the start of fiscal 2027, which cut its reported 2026 capex from about $190 billion to about $175 billion. Alphabet's 2026 guidance comes from coverage of its earnings call. For reference, Amazon spent $83.0 billion in 2024, Alphabet's first nine months of 2026 came to $132.0 billion, Amazon's to $175.7 billion and Meta's to $83.1 billion. Meta narrowed its 2026 outlook from $130–145 billion.

Alphabet spent $32.3 billion in 2023, $52.5 billion in 2024 and $91.4 billion in 2025, then guided to $195 to $205 billion for 2026. Amazon spent $52.7 billion in 2023 and $131.8 billion in 2025, and expected about $220 billion in 2026. Meta spent $27.0 billion in 2023 and $69.7 billion in 2025, and guided to $135 to $145 billion in 2026. Microsoft, whose fiscal year ends in June, went from $28.1 billion in fiscal 2023 to $116.0 billion in fiscal 2026, and guided to $170 billion, including finance leases, for calendar 2027.

Add up the 2026 guidance of the first three alone and you get about $550 to $570 billion in a single year. Peak to peak, Meta and Alphabet each multiplied their annual spending about five- or sixfold in three years. Amazon alone spent $131.8 billion in 2025. The carriers took three years to spend their hundred billion.

Paying for It

For most of the cloud era, these were the best cash machines in business history. They could afford almost anything. In 2026, the spending finally ran past the cash.

Amazon's trailing free cash flow, the cash left after paying for its buildings and equipment, was an outflow of $42.4 billion through September 2026. Its filing explained why: "a year-over-year increase of $94.8 billion in purchases of property and equipment … primarily reflects investments in artificial intelligence." Alphabet's free cash flow was negative $9.1 billion in the third quarter, after negative $5.9 billion in the second. Its chief financial officer, Anat Ashkenazi, said 2027 capex would "increase significantly."

Investors didn't always cheer. When Amazon guided to about $200 billion of 2026 spending in February 2026, against Wall Street's estimate of $146.6 billion, its shares fell more than 10 percent after hours. Andy Jassy didn't flinch: "we're monetizing capacity as fast as we can install it." By October, he had raised the number to about $220 billion, and his ambition with it: "There has never been a trillion dollar revenue business, and we may have two of them in the next 5-10 years." AWS revenue that quarter grew 39 percent, to $46.0 billion.

Manhattan in Louisiana

Mark Zuckerberg made the scale visual. In July 2025 he announced Prometheus, a 1-gigawatt cluster due online in 2026, and Hyperion, in Richland Parish, Louisiana. "We're also building Hyperion, which will be able to scale up to 5GW over several years," he wrote. "Just one of these covers a significant part of the footprint of Manhattan."

When Meta announced Hyperion in December 2024, it was estimated to cost $10 billion. Ten months later, Bloomberg reported how Meta would pay for a much bigger version: about $27 billion of debt and $1.5 billion of equity raised through a vehicle run by Blue Owl Capital and arranged by Morgan Stanley. Meta kept 20 percent. The debt matures in 2049. And the structure kept it off Meta's balance sheet.S&P figures cited elsewhere put the commitments at $28.79 billion, with Blue Owl at $23.03 billion and Meta at $5.76 billion. Hyperion is now expected to exceed 5 gigawatts and come online in 2029. A July 2026 report that it is expanding to more than $50 billion comes from a secondary source and is not verified.

Off-balance-sheet financing is a phrase that makes anyone who lived through 2001 sit up straight. To be fair, Meta disclosed the deal, Blue Owl's investors knew exactly what they were buying, and a 2049 bond backed by a lease to Meta is a different animal from a fiber swap. But it tells you something that the richest companies in the world had started borrowing other people's balance sheets to build.

Hyperion also needed power. By March 2026, Meta had ordered ten gas plants for the campus from the utility Entergy, totaling 7.5 gigawatts, "a more than 30% increase to Louisiana's entire grid capacity," plus 2.5 gigawatts of renewables. Meta agreed to pay $11 billion, and the site grew to 3,650 acres. Meta's Rachel Peterson said the company was paying its own way "so that Entergy's other consumers aren't paying our costs."

That sentence would become the argument of the next two years. Before we get to it, we need to talk about where the power came from.

Power—Joe Dominguez#

In the fiber Boom, the scarce thing was capital, and the abundant thing was fiber. In the AI Boom, the scarce thing is electricity.

You can watch the change in the companies' own risk warnings. In its 2004 annual report, Google treated electricity as a reliability risk, a reason to worry about backup generators and fuel in an outage. Microsoft's 2022 report said its data centers "depend on predictable energy." By February 2026, Alphabet's annual report said its "ability to scale our technical infrastructure is increasingly constrained by the availability of power, water, and land." Electricity had gone from a backup risk to the limit on growth.

In December 2024, Lawrence Berkeley National Laboratory reported that U.S. data centers used about 76 terawatt-hours of electricity in 2018, 1.9 percent of the nation's total. By 2023, they used 176 terawatt-hours, "representing 4.4% of total U.S. electricity consumption." AI hardware, the lab wrote, "has ended the era of generally flat data center energy use." It projected 325 to 580 terawatt-hours by 2028, or 6.7 to 12.0 percent of the country's power, the equivalent of 74 to 132 gigawatts of demand.

The International Energy Agency's Energy and AI report, in April 2025, put the world's data centers at 415 terawatt-hours in 2024, about 1.5 percent of global electricity. Its base case had them at about 945 terawatt-hours by 2030, slightly more than all of Japan uses today. In the United States, it said, data centers were on course to account for almost half of the growth in electricity demand. The agency also said that about 20 percent of planned data-center projects could be delayed by grid constraints. "AI is one of the biggest stories in the energy world today," said Fatih Birol, its executive director.

Those are forecasts, and this book has a long memory about forecasts. But the utilities treated them as orders.

Three Mile Island

On September 20, 2024, Constellation Energy announced that it would restart Unit 1 at Three Mile Island, in Pennsylvania, and sell all of its roughly 835 megawatts to Microsoft for twenty years. The reactor had shut down in 2019 for economic reasons. It had never been the reactor that melted down in 1979; that was Unit 2. But the name carried the weight of America's worst commercial nuclear accident, so Constellation renamed the plant the Crane Clean Energy Center, for the late Exelon CEO Chris Crane.

Constellation's CEO, Joe Dominguez, made the case: "nuclear plants are the only energy sources that can consistently deliver on that promise." Microsoft's Bobby Hollis called it "a major milestone in Microsoft's efforts to help decarbonize the grid." The original target was 2028. Constellation later moved it up to 2027, won a federal waiver in June 2026 to transfer 760 megawatts of grid-connection rights from the Eddystone plant to Crane, and landed a $1 billion Department of Energy loan. As of mid-October 2026, the Nuclear Regulatory Commission's final approval had not been confirmed, and a PJM analysis suggested full grid connection might not come until about 2031.The Nuclear Regulatory Commission issued a draft environmental assessment on June 8, 2026. Sources conflict on the status of final approval, and the 2027 restart target conflicts with the PJM analysis suggesting full transmission access around 2031.

Think about what that deal says. A software company signed a twenty-year contract that brought a shuttered nuclear reactor back to life. In Bandwidth, the customers bought capacity on our networks. Here, the customer bought the power plant's whole output.

Talen and the Grid

Amazon tried to go one step further and plug a data center directly into a nuclear plant. Talen Energy's Susquehanna plant in Pennsylvania produces 2,494 megawatts. Amazon bought a data-center campus next door and asked to take power "behind the meter," without passing through the grid. Late on November 1, 2024, the Federal Energy Regulatory Commission voted 2–1 to reject a revised agreement that would have raised that direct load from 300 to 480 megawatts. Two big utilities, Exelon and AEP, had protested that the arrangement could shift as much as $140 million a year of costs onto everyone else on the grid. The majority found that PJM had not met "the high burden" of justifying the special terms. Commissioner Mark Christie joined "because PJM has failed to meet its burden of proof on this record," and warned that co-location raised "complicated, nuanced and multifaceted issues." The chairman, Willie Phillips, dissented. He called the order "a step backward for both electric reliability and national security," and said "access to reliable electricity is the lifeblood of those data centers." The question underneath was simple: if a data center takes power from a plant without using the grid, who pays for the grid?

Amazon and Talen restructured. In June 2025, they signed a deal for up to 1,920 megawatts through 2042, delivered through the grid this time, ramping up to full volume by 2032.FERC's vote was Christie and See to reject, with Chairman Phillips dissenting; Rosner and Chang did not participate. Under the restructured contract, deliveries ramp to 840 to 1,200 megawatts by 2029. Power Magazine valued the deal at about $18 billion over seventeen years; that figure has a single source.

The tech companies also placed bets on reactors that don't exist yet. In October 2024, Google agreed to buy up to 500 megawatts from six or seven small modular reactors to be built by Kairos Power, the first by 2030. Two days later, Amazon led a roughly $500 million investment in X-energy, which aimed to build more than 5 gigawatts of small reactors by 2039, and signed agreements with Energy Northwest in Washington State and Dominion in Virginia. These are the most hopeful deals in this chapter and the furthest from delivering a single electron.

Gas, Again

Even Data Center Alley ran short. In July 2025, Jeff Tench of Vantage Data Centers, which he described as "a Colorado-based developer," told a Senate committee, "Five years ago, a 30-megawatt facility was considered large. Today, 100 megawatts is the starting point." The local utility had planned to give Vantage 100 megawatts for a Loudoun County project. After construction began, Vantage was told the power "would not be available for five to seven years." The site sat on a high-pressure gas line, so Vantage built its own power plant, independent of the grid, in eighteen months. "We were left no choice," Tench said.

What the industry could build quickly was gas. In Memphis, Southaven, Louisiana and Piketon, the answer to "where will the power come from?" was the same. By the end of 2025, according to Lawrence Berkeley's annual survey, gas projects waiting to connect to the U.S. grid had jumped 86 percent in a year, to 253 gigawatts, while solar, storage and wind projects fell. GE Vernova's CEO, Scott Strazik, said in December 2025 that the company's gas turbines were sold out through 2028.

Even gas plants have to wait their turn. Lawrence Berkeley found that the median power project completed in 2025 took sixty-one months from its request to connect to the grid until it started operating. In 2015 it took thirty-six months. In 2008, it took twenty-two. About 2,061 gigawatts of projects were in the queue at the end of 2025. Only about 13 percent of the capacity requested from 2000 to 2020 ever got built.

PJM

The place the shortage showed up first was PJM, the grid operator for thirteen states and Washington, D.C., including Northern Virginia's Data Center Alley. PJM runs an annual auction to pay power plants to be available three years out. The price tells you how tight the grid is.

For the 2024–25 delivery year, PJM's capacity price was $28.92 per megawatt-day. For 2025–26, it jumped to $269.92, more than nine times higher. For 2026–27, it hit $329.17, a cap set under a settlement with Pennsylvania's governor, Josh Shapiro. The December 2025 auction, for 2027–28, cleared at the cap of $333.44, cost $16.4 billion and, for the first time, left PJM 6,623 megawatts short of what it needed for reliability. PJM attributed nearly 5,100 megawatts of the 5,250-megawatt increase in forecast demand to data centers.PJM itself does not attribute the price increases to data centers. Advocacy analyses (NRDC and Synapse) attribute about 63 percent, or $9.3 billion, of recent capacity costs to data centers, and the Citizens Utility Board calls large loads "the biggest contributor." The 2027–28 cleared mix was 43 percent gas, 21 percent nuclear and 20 percent coal, and only 774 megawatts of new generation and uprates cleared. In the 2028–29 auction, the uncapped price in the ComEd zone would have been $776.69. Stu Bresler, a PJM executive, didn't mince words: "this auction leaves no doubt that data centers' demand for electricity continues to far outstrip new supply."

The July 2026 auction, for 2028–29, came up short again, by about 6,831 megawatts, the third shortfall in a row. Without the cap, the price would have been $554.72. Only 525 megawatts of new resources cleared.

Texas took a different approach. In June 2025, Governor Greg Abbott signed a law, Senate Bill 6, that lets ERCOT, the state's grid operator, order any load of 75 megawatts or more to switch to backup power or shut down in an emergency.Sources differ on whether Abbott signed SB 6 on June 20 or June 21, 2025. It became known as the kill switch.

Peak to trough is my usual move in this book. Here it runs backward: PJM's capacity price rose more than elevenfold in four auctions. In the fiber Boom, the glut showed up in falling prices. In this one, the scarcity is showing up in rising ones. And the bill lands on everyone connected to the grid, households included.

Resources and Backlash#

Every Boom needs the community's consent, even if it doesn't ask for it. In the AI Boom, the fights were over water, power bills and what kind of neighbor a gigawatt makes. Both sides had real arguments.

The Case Against

Start with water. Many data centers cool their chips by evaporating water. Lawrence Berkeley estimated that U.S. data centers used 66 billion liters of water directly in 2023, about 17.4 billion gallons, up from 21.2 billion liters in 2014. That doesn't count the much larger volume used to generate their electricity.

The Associated Press brought it home in 2023. Microsoft's water use rose 34 percent from 2021 to 2022, to nearly 1.7 billion gallons. In July 2022, its campus in West Des Moines, Iowa, built to train OpenAI's GPT-4, drew about 11.5 million gallons, about 6 percent of the local water district's supply. "It's fair to say the majority of the growth is due to AI," said Shaolei Ren, a researcher who studies the problem. Google reported using 7.2 billion gallons of fresh water in 2024 and replenishing about 64 percent of it.Other summaries of Google's report give about 6 billion gallons; the figure depends on how fresh water is defined. Google's Council Bluffs, Iowa campus used about 1 billion gallons. Ren estimates about 500 milliliters of water per 5 to 50 ChatGPT prompts, counting indirect use.

In Tucson, Arizona, in August 2025, the city council voted 7–0 to end the annexation that a data-center project called Project Blue needed to get city water. The developer, Beale Infrastructure, called the vote "a missed opportunity" and said the project would have been "net water positive." Critics also objected to the nondisclosure agreements wrapped around the project.

Then there were the power bills. In December 2024, Virginia's legislative watchdog, the Joint Legislative Audit and Review Commission, or JLARC, published the most careful public study of the question. It called Northern Virginia "the largest data center market in the world," with about 13 percent of global capacity. Virginia's power demand had been "essentially flat from 2006 to 2020." Unconstrained, it "would double within the next 10 years," mainly because of data centers. Meeting that would take a large gas plant almost every year for fifteen years, or about every year and a half under the state's clean-energy law. JLARC found that current rates "appropriately allocate costs," so data centers were paying their own way today. But it estimated that generation and transmission costs for a typical Dominion household could rise $14 to $37 a month by 2040, in 2024 dollars.Joint Legislative Audit and Review Commission, Data Centers in Virginia, Report 598, December 9, 2024. The $14–$37 range is for generation and transmission on a typical $90 monthly charge, in constant 2024 dollars. Under unconstrained demand the increase is $22–$23 by 2030 and $33–$37 by 2040; at half that demand, $6–$7 by 2030 and $14 by 2040. Some coverage attributed to JLARC a need for "150 percent more in-state generation" and bills "nearly $40" a month higher, along with a quote from its project leader, Mark Gribbin. None of those appears in the report, and I have used the report. JLARC counted about 150 data-center sites in Virginia using about 5,050 megawatts, roughly the electricity needs of 60 percent of the state's households. Residential electricity rates nationally rose 6 percent in 2025, more than twice the rate of inflation.

The voters noticed. A Heatmap poll found that 71 percent of Americans would oppose a data center near their home. In Prince William County, Virginia, the Digital Gateway, a giant data-center project near the Manassas battlefield with room for up to thirty-seven buildings, had passed 4–3 with one abstention in December 2023, after more than seventeen hours of public comment at a meeting that ran twenty-seven hours. It died over a newspaper ad. The county clerk asked the Washington Post to run the legally required notice. The Post quoted $7,509.48 and asked for confirmation by 3 p.m. the next day. In the words of the Court of Appeals of Virginia, "The clerk failed to call or respond by the deadline." A judge voided the rezoning in August 2025. The appeals court affirmed in March 2026, and in July 2026, QTS withdrew its last appeal. The project was dead, though nothing barred new rezoning applications.

The Case For

The industry's arguments were also real, and the best of them came from the same places.

Loudoun County, Virginia, home of Data Center Alley, has more data centers than almost anywhere on earth. Its Chamber of Commerce says data centers generated "nearly $1.2 billion in County revenues" and helped cut the property tax rate from $1.245 to 80.5 cents. JLARC put data centers at 31 percent of Loudoun's total local tax revenue.The Chamber's figure of about 38 percent uses county funding as its base; JLARC's 31 percent is of total local tax revenue. In Prince William County, JLARC put the share at 7 percent. Data Center Frontier reported in 2022 that data centers were expected to pay $575 million of Loudoun's taxes that year, about 31 percent of county revenue. It found that the industry supported about 74,000 jobs, $5.5 billion of labor income and $9.1 billion of Virginia's GDP a year, though most of that came from construction; a typical finished data center employs about fifty people. JLARC rated the state's sales-tax exemption for data-center equipment, worth $928 million in fiscal 2023, as a "moderate" return that, like most incentives, "does not pay for itself." Its recommendation was to let utilities "delay, but not deny" service to big new loads. Josh Levi of the Data Center Coalition said the study "validates Virginia's leadership." The industry's own study, by PwC, counted 603,900 direct jobs in 2023 and about $163 billion in taxes, though critics at the Ohio River Valley Institute argued that the multipliers were inflated.The PwC study counted 4.7 million total jobs and $162.7 to $164.7 billion of taxes; a 2024 update cites $926.9 billion of GDP and $204.4 billion of taxes. The Ohio River Valley Institute argues that direct state and local taxes in Pennsylvania were only $1.36 billion.

On power bills, the evidence was more mixed than the campaign ads suggested. A preprint from EPRI, the utility industry's research institute, found that states with data-center growth saw retail rates about 6 percent lower from 2019 to 2024, because the data centers helped pay the utilities' fixed costs. Lawrence Berkeley found something similar. A Google-supported synthesis from Columbia University's energy center agreed that the benefit was possible, with a warning. "This outcome is not automatic," said Robin Millican. "It depends on supply availability, cost allocation, and tariff design."

Tariff design was the industry's best answer. On July 9, 2025, Ohio's utility commission voted 5–0 to approve a tariff for AEP Ohio requiring new data centers of 25 megawatts or more to pay for at least 85 percent of the capacity they reserve, for up to twelve years, whether they use it or not. A customer without an A-minus credit rating and enough cash had to post collateral equal to half of its total minimum charges. In return, the commission ordered AEP Ohio to lift its moratorium on new data-center hookups "as soon as possible." The commission's chair, Jenifer French, called it "a well-balanced package that safeguards non-data center customers." The data-center companies had filed a competing settlement and lost.Public Utilities Commission of Ohio, Case No. 24-508-EL-ATA. The vote is from the commission's minutes of July 9, 2025, and the tariff terms from AEP Ohio's published Schedule DCT; I have not read the written opinion and order. Minimum billed demand ramps over up to four years (50, 65, 80 and 90 percent) before an eight-year term. A customer that exits after the fifth year past the ramp pays thirty-six months of minimum charges. The data-center companies' settlement was filed October 10, 2024, and AEP Ohio's with the commission staff and the Consumers' Counsel on October 23. The tariff took effect July 23, 2025. Meta's Peterson was making the same argument in Louisiana. If the data centers pay their own way, the neighbors don't pay for them.

By 2026 the model labs were saying it too. Anthropic's IPO filing told investors, "We have publicly committed to covering the cost of electricity price increases to consumers caused by our data centers." The same filing warned that "Opposition to data center development has also become more organized and coordinated across jurisdictions."

Even Loudoun changed its mind, partway. In March 2025, its supervisors voted 7–2 to end "by-right" approval for new data centers, so that every new project would need a public vote. Twenty-two pending applications were allowed to continue. The county most dependent on data centers had decided to keep them and to have a say.

Texas Turns

Nowhere did the politics move faster than in Texas. Greg Abbott had promoted Texas as "the epicenter of AI development" in November 2025. Abilene, Shackelford County and Milam County were all Stargate sites. Then came the strain on the grid and the water. In June 2026, Abbott ordered data centers to pay for their own grid infrastructure. On August 3, he paused new data-center connections to ERCOT. On September 21, he went further and ordered the state's environmental agency to stop issuing permits until an audit of the grid and water supply was done. As few as 28 percent of data centers had answered a state survey on their water use. "Until they do," Abbott said, "TCEQ will issue no permits sought by data center projects."

BloombergNEF estimated that nearly 20 percent of the nation's 253-gigawatt data-center pipeline, about 50 gigawatts, could be at risk from the Texas pause. Data Center Watch, which tracks local opposition, claimed that $68 billion of projects had been delayed or canceled by local pushback in a single quarter.

Washington pulled the other way. In July 2025, President Trump's Executive Order 14318 sped up federal permitting for projects with more than 100 megawatts of new AI load, opened federal lands to data centers and told the EPA to expedite Clean Air Act permits. It didn't override the states. And the states, it turned out, were where the voters were.

November 3, 2026

The midterm elections were three days ago as I write this, and some results are still unofficial.Results here come from AP, CNBC and MIT Technology Review, November 4–6, 2026. Confirm against certified results before print. Here is what the sources report.

Ballotpedia counted forty-one local ballot measures on data centers in ten states, twenty-one of them in Ohio. Voters went against data centers in every one except Harwood, North Dakota. AdImpact counted more than $380 million of TV ads that mentioned data centers, with Democrats outspending Republicans more than 3 to 1. In the CNBC Voter Poll, 56 percent of voters called data centers a crucial or major factor in their vote.

Democrats flipped governorships in Texas, Ohio, Georgia, Nevada and Iowa, many of them data-center hot spots. In Texas, Gina Hinojosa beat Abbott. She had supported data-center tax breaks in 2023 and campaigned in 2026 for a "full-fledged moratorium." James Talarico beat Ken Paxton for the Senate, and Democrats won the Texas House for the first time in twenty-four years. Musk spent at least $474 million and lost both of the marquee Texas races. In Ohio, Amy Acton, who backed a "conditional moratorium," beat Vivek Ramaswamy for governor, and Sherrod Brown beat Senator Jon Husted. A year earlier, in November 2025, Georgia voters had already flipped two seats on the state's utility commission to Democrats, who ran against soaring power bills.

"The revolt against data centers was one of the top issues of the 2026 election," said Representative Greg Casar. The industry warned of "real world consequences," in the words of Dan Diorio of the Data Center Coalition. OpenAI asked for "clear rules for how the industry builds."

What it means is still unclear. MIT Technology Review noted that in Texas, "no single office in the state decides whether a project goes ahead or is halted." But the projects at stake were enormous: Stargate Frontier in Shackelford County, at 1,820 megawatts, and a campus called Nexus Hubbard, at 6,400 megawatts, mostly for Anthropic.

In Bandwidth, the people who paid for the Bust were investors. This time, the voters were trying to make sure it wouldn't be them.

The Bubble Debate—Odlyzko's Ghost#

On Thursday, November 5, 2026, two days after the election, the Nasdaq fell 5.52 percent, its worst day in nineteen months. The S&P 500 fell 4.36 percent and lost about $3 trillion of value. The index of chip stocks fell more than 12 percent. Nvidia fell more than 9 percent.Investing.com reported that the Nasdaq 100 fell 6.4 percent, "its largest single-day decline on record," and other outlets said it was the worst day since 2020. Reuters called it the Nasdaq Composite's worst day in nineteen months. I've used Reuters. The VIX volatility index hit its highest level since April 2025. The Nasdaq closed more than 10 percent below its October 6 record, in a correction. Reuters blamed fears that the new Democratic majorities would slow the build-out. "Some newly elected members of Congress and governors are looking to slow the pace of AI infrastructure development," said Melissa Brown of SimCorp.

One bad day isn't a Bust. But it's a good day to ask the question this whole chapter has been building toward. Is this a bubble?

The surprising answer is that the man at the center of it said yes. At a dinner with reporters in August 2025, Sam Altman was asked whether investors were overexcited about AI. "My opinion is yes," he said. He compared it to the dot-com bubble, when, he said, "there was a real thing" underneath the excess. In the same conversation he said, "You should expect OpenAI to spend trillions of dollars on data center construction in the not very distant future."

That's the whole debate in two sentences from one man. The thing is real. The spending is enormous. Both can be true, and in my experience, both usually are.

The Traffic Myth

This is where the last Boom earns its place in this book.

As I wrote in Bandwidth, in 1998 an AT&T Labs researcher named Andrew Odlyzko published a paper debunking the most important number of the Bandwidth Boom: the belief that Internet traffic was doubling every three months, or every hundred days. Odlyzko pointed out that if that were true, traffic by 2001 would have grown by a factor of seventeen million. It would have required every Internet user on Earth to stream video twenty-four hours a day. His measurements showed traffic doubling about once a year. That was still spectacular growth. It just wasn't enough to fill the networks we were building.

The myth had started as a true number. In a 2001 paper, Odlyzko and his AT&T colleague K. G. Coffman showed that U.S. traffic really had doubled every three to four months in 1995 and 1996. By 1997 it had slowed to about doubling each year, and it stayed there. The industry kept quoting the old rate for four more years. "The telecom crash of 2000 occurred," they wrote, "when investors realized that transmission capacity in place and under construction greatly exceeded actual traffic demand."

"[Doubling every three months] was an extremely convenient myth," Odlyzko said. "Every entrepreneur who was getting financing could quote it."

And the analysts amplified it. Jack Grubman, the highest-profile telecom analyst on Wall Street, proclaimed that bandwidth requirements were doubling every four months. Grubman's word moved billions. "I'm sculpting the industry," he said. His firm collected more than $1 billion in fees from the companies he covered, and when he was asked about the conflict, he answered with his trademark swagger: "What used to be a conflict is now a synergy." And the whole analyst community found a way to value companies that had no earnings. As I quoted in Bandwidth, the New York Times explained the method in its article "Pied Piper": analysts valued the companies "based on the money they had put into their networks—money from investors, not money that the company had earned." Telecom shares were promoted as good values "because they traded at prices that represented four or even five times what they had invested in plant and equipment." Hold that thought. We will need it when we get to today's valuations.

Odlyzko's inconvenient truth, I wrote in Bandwidth, "was either unnoticed or, more likely, ignored."

Today's Doubling Claims

So what is today's version of "doubling every hundred days"?

The AI industry's version comes in three forms. The first is the power forecast: Lawrence Berkeley's range of 325 to 580 terawatt-hours by 2028, a spread so wide that the high end is nearly double the low end.

Power forecasts have a history of their own, and one witness spans both booms. In February 2000, Mark Mills told a House subcommittee that the Internet was swamping the grid. Another witness, Joseph Romm, answered, "Mills is wrong by at least a factor of 8," citing Lawrence Berkeley's scientists. Romm was right. In April 2025, Mills was back before the same subcommittee's successor, warning that new data centers would need "somewhere between 70 gigawatts and 130 gigawatts more power capacity" than planners had expected two years earlier. This time Lawrence Berkeley's own numbers showed a real surge. But another witness, Tyson Slocum of Public Citizen, reached for the telecom era by name. "Forecasters consistently overestimate electricity demand," he said. The Internet power boom forecast after the 1996 Telecom Act, he said, "never materialized."

The second is the backlog. In the Exodus section of Chapter III, I wrote that "in the Gold Rush, the forecast was the asset." In this Gold Rush, the asset is the contract. Oracle reports $664 billion of remaining performance obligations. CoreWeave reports a backlog of more than $100 billion. Those numbers are signed, which makes them much better than a forecast. But a contract is only as good as the customer's ability to pay, and the largest customers in this chapter, OpenAI and Anthropic, were raising new money every few months to pay for what they had signed.

The third claim is the circle. Follow the money in a few of the deals in this chapter. In September 2025, Nvidia agreed to invest up to $100 billion in OpenAI, $10 billion at signing and the rest as OpenAI deployed at least 10 gigawatts of Nvidia systems.Kress later said the Nvidia–OpenAI agreement was not finalized. By March 2026, Nvidia had put $30 billion into OpenAI's funding round. The first gigawatt was planned for the second half of 2026 on Vera Rubin. On the AMD deal, at the $600 share-price target for the final tranche, OpenAI's shares would be worth $96 billion. Stacy Rasgon of Bernstein said the deal "will fuel" concerns about circularity. Two weeks later, OpenAI agreed to deploy 6 gigawatts of AMD's chips, and AMD gave OpenAI a warrant to buy up to 160 million AMD shares, about 10 percent of the company, for a penny each as the chips were deployed and AMD's stock price hit its targets. Amazon invested in Anthropic, which committed to buy at least $110 billion of Amazon's cloud services. Microsoft and Nvidia invested in Anthropic too, while selling it computing.

Readers of Bandwidth will remember round-tripping, the capacity swaps that Global Crossing and Qwest used to inflate each other's revenue. Let me be very clear about the difference. Those swaps, as I wrote in Bandwidth, let both parties pump up their financials, and at Qwest they were part of a fraud. Nothing in the sources for this chapter suggests anything of the kind today. These deals are disclosed, the chips are real and the computing is used. But a vendor that finances its own customers is taking a bet on those customers, and the revenue it reports is only as solid as their next funding round. That is a fair question to ask.

The Skeptics

The skeptics had been asking for three years, and their best weapon was arithmetic.

In September 2023, David Cahn of Sequoia Capital, a venture firm, posed "AI's $200B Question." His method was simple. For every dollar spent on Nvidia's chips, he reckoned, someone spends another dollar on the building and the power around them. Whoever rents out the chips needs a profit margin on top. Add up the AI revenue anyone could plausibly point to, and he found a gap of more than $125 billion a year. Nine months later, in June 2024, he ran the numbers again. He doubled Nvidia's data-center revenue run-rate once for the buildings and power, and doubled it again for the renter's 50 percent gross margin, and got $600 billion a year of AI revenue needed to earn back the spending. He called it "AI's $600B Question." The hole had grown to about $500 billion. Cahn wasn't a bear. "Speculative frenzies are part of technology, and so they are not something to be afraid of," he wrote. But he had a line for a book about fiber: "It's hard to pick winners, but much easier to pick losers (canals, in the case of railroads)."

The same month, Goldman Sachs published an issue of its Top of Mind series titled "Gen AI: Too Much Spend, Too Little Benefit?" It was a debate inside the firm. Jim Covello, head of global equity research, was the skeptic. He estimated the build-out would cost more than $1 trillion and asked, "What $1tn problem will AI solve?"Goldman's page was blocked to automated reading. Covello's "$1tn" question is quoted from The Daily Upside's coverage, his "exceptionally expensive" line from 404 Media's, and the Sun server comparison and Nvidia quote from Contrary Research's summary of the issue. Rangan's view is paraphrased from press summaries. In the same issue, MIT's Daron Acemoglu estimated that AI would affect less than 5 percent of tasks within a decade, while other Goldman economists were more optimistic. "AI technology is exceptionally expensive," he said, "and to justify those costs, the technology must be able to solve complex problems." He rejected the idea that the costs would fall the way the Internet's had. The Internet of the late 1990s, he pointed out, ran on Sun servers that cost $64,000. "Today, Nvidia is the only company currently capable of producing the GPUs that power AI." Readers of Chapter III will remember those Sun servers. In the same issue, Kash Rangan took the other side, arguing that the AI cycle was still in its build-out phase and that the killer application would take time to find.

The consultants and the banks kept enlarging the hole. In September 2025, Bain & Company estimated that by 2030 the industry would need about $2 trillion of annual revenue to pay for the computing it was building, and that even after AI-driven savings it would come up about $800 billion a year short. In November 2025, J.P. Morgan's researchers calculated that a 10 percent return on the AI build-out required about $650 billion of revenue a year, which they translated as about $35 a month from every iPhone user. Morgan Stanley put global data-center spending from 2025 through 2028 at about $2.9 trillion, of which the cloud companies could pay for about $1.4 trillion out of their own cash flow. The rest, a financing gap of about $1.5 trillion, would have to be borrowed, much of it from private credit.Bain & Company, press release for its sixth Global Technology Report, September 23, 2025; J.P. Morgan Research, "AI Capex – Financing the Investment Cycle," November 10, 2025, as summarized by J.P. Morgan Asset Management, January 9, 2026; Morgan Stanley's figures as described in its Thoughts on the Market podcast of August 6, 2025, which refers to a July 2025 research note I have not read. Cahn's 2023 estimate assumed a $50 billion Nvidia run-rate; his 2024 estimate implied about $150 billion.

In August 2025, a study from MIT found that about 95 percent of organizations saw no measurable return on their generative AI spending, though "no measurable return" meant no measurable profit impact yet, not zero value.

Then came Michael Burry, who had famously bet against the housing market before 2008. In November 2025, Burry argued that the cloud companies were understating depreciation by about $176 billion over 2026 through 2028, by assuming their GPUs last about six years when, he said, two or three was more realistic. Stretch the life of an asset and you shrink the annual cost, which raises reported earnings. Burry was short AI stocks when he said it. The bulls answered with the evidence I mentioned in the CoreWeave section: five-year-old chips still earning comfortable margins, and three-year-old chips re-leased at nearly the old price.

That is the modern form of a very old question. In Bandwidth, the companies that failed were the ones that confused building with earning. Capital spending made the networks bigger, but only cash flow made the companies worth anything. Today's question is whether the depreciation schedules, the backlogs and the circular investments are measuring cash flow, or just measuring construction.

The Bulls

The bulls had a better answer in 2026 than anyone had in 2000. The revenue is real. Azure grew 48 percent in a quarter. AWS grew 39 percent. Google Cloud's revenue reportedly grew 103 percent. Anthropic's revenue went from $41 million in 2023 to $33.2 billion in the first nine months of 2026. The fiber companies of 2000 never had a customer base growing like that. And every cloud executive in this chapter said the same thing: demand exceeds supply. Jassy said Amazon was monetizing capacity as fast as it could install it. Hood said Microsoft still had more demand than supply. Huang said compute is revenue.

Aswath Damodaran, the NYU finance professor who has spent his career valuing companies by their cash flows, put the Anthropic IPO to that test in October 2026. He assumed Anthropic would reach about $1 trillion of revenue by 2036, at a 27.5 percent margin, with a 10 percent chance of failure, and came out at about $1.475 trillion for the equity. That was at the low end of the $1.5 to $2 trillion being whispered at the time, and he said he would sit out if it priced well above. It was a cash-flow valuation of a company that was still losing money. And it still came out above a trillion.

The Analyst Chorus

In Bandwidth, I told how MoffettNathanson's Nick Del Deo greeted Zayo's IPO in October 2014 with "Batman vs. Superman," a report asking whether dark fiber deserved the valuation of a tower. Eleven years later, Del Deo covered the data-center landlords and CoreWeave, and his notes from the AI Boom read like a two-sided chorus by themselves.

The bull side was demand. In February 2024, in a note titled "Bulls vs. Bears in a Time of Record Demand," he started out expecting to downgrade Digital Realty and Equinix and ended up keeping his ratings. "There's no reason to believe demand will suddenly roll over," he wrote. Scarcity, he later found, had given the landlords something they hadn't had in the cloud era: pricing power. And in December 2025, measuring the two big landlords by value per megawatt, he found them oddly cheap. Digital Realty's stock had risen 24 percent since the end of 2023 and Equinix's had fallen 2 percent, while the S&P 500 rose 48 percent. Their valuation multiples, he wrote, "certainly do not suggest that there's currently a data center boom taking place."

The bear side was everything around the demand. When DeepSeek's cheaper model rattled the market in early 2025, the question was whether the Jevons paradox would hold: make a resource cheaper to use and people use more of it. Del Deo, in a note he admitted "may be the clunkiest title we've ever used," pointed to an industry I know well. Internet transit had looked like an obvious Jevons case, and it didn't play out that way: traffic soared, and falling prices ate the gains. "Examples exist of seemingly obvious candidates for Jevons Paradox not playing out as expected," he wrote. More efficient AI could raise volumes and still shrink the landlords' value creation. And if AI was a bubble, he argued in his October 2025 telecom comparison, the reckoning would come the way it did in telecom, from monetization: the revenue that never arrived.

His house view, in short, was real demand, fragile financing, and the old landlords safer than the upstarts.

The historian had his say too. Andrew Odlyzko, the man who debunked the doubling myth, spoke publicly about the AI Boom in the fall of 2025, and he didn't sound like a man repeating himself. He drew a distinction that matters for this book. "The dotcom bubble did not involve very much real investment," he told Transformer's James Ball. Telecom did, and those investments "were based on the assumption that internet traffic was growing tenfold per year, and it wasn't." The better comparison, he said, was older: "The great British railway mania of the 1840s was the largest technology bubble in history." The railways got used. "Hardly any of the lines that were built were left unused." But they were built, he said, "on the bodies of the investors."

What worried him this time was the money, not the demand. Creative finance, he told the Bulletin of the Atomic Scientists in December 2025, "generates these very complex structures which nobody fully understands until things blow up." He added, "I'm seeing more of that, and that's what is getting me concerned; this part is typical of bubbles."

The Questions

I won't predict how this ends. Buffett's clock has no hands, and I'm not going to pretend I can see it. But I can ask the questions I wish someone had asked loudly in 1999.

Who is today's Odlyzko? Odlyzko himself is still asking about the financing. But who is measuring the thing itself, tokens actually used and paid for, rather than the forecasts of the people selling the capacity? The closest thing may be a regulator's discount. In 2025, Xcel Energy had 5.8 gigawatts of data-center applications pending in Colorado, about as much as its entire 6.2 gigawatts of generation in the state. A Colorado utility commissioner, Megan Gilman, called it "a really massive load forecast far in excess of what we've seen on the system," and said "the vast majority of that forecast is far from certain." Nobody discounted the traffic forecasts that way in 1999.

Who are today's Grubmans? Twenty-two investment banks are underwriting the Anthropic IPO. Which of them will publish a sell rating on it?

What happens to the backlog if one large customer can't raise its next round? Who is holding the contract, and who is holding the debt?

And what is a GPU worth in three years? If the answer is "most of what it cost," the build-out is an investment. If the answer is "not much," it's an expense that's being called an investment.

In 2001, the fiber was still in the ground after the companies that built it were gone, and the Phoenix companies of the next decade, Zayo among them, built businesses on it. This time, which part of the build-out is the fiber in the ground, and which part will be obsolete before it's paid for? Odlyzko's railways and Del Deo's conduit point to the same answer: the long-lived assets, the powered land, the buildings and the transmission lines, are the likeliest to outlast their first owners. The chips are the likeliest to be repriced.

Anthropic—Dario and Daniela Amodei#

A disclosure belongs here, in plain sight. This book is researched and drafted with Anthropic's Claude. It is an experiment with Claude. So the section you are about to read, about Anthropic, was drafted in part by Anthropic's product. As Grubman would say, what used to be a conflict is now a synergy. I have tried to hold it to a stricter standard than Grubman held himself: every number below comes from Anthropic's SEC filing or is labeled as reporting.

Anthropic was founded in 2021 by a group of researchers who left OpenAI, led by Dario and Daniela Amodei, who are brother and sister.The S-1 says Anthropic was "founded in 2021 by a team of researchers"; its "Founder LLC" holds the Amodeis and five other co-founders. Dealroom counts seven ex-OpenAI founders; some sources say more. The S-1's January 2021 incorporation date settles earlier conflicts over the founding date. Dario is the chief executive. Daniela is the president and chair of the board. The company was "incorporated as a Delaware PBC in January 2021," a public benefit corporation, which means Delaware law requires it to weigh the public benefit named in its charter alongside profit. Its $124 million Series A, announced in May 2021, was led by Jaan Tallinn, a co-founder of Skype, with Dustin Moskovitz participating.

Anthropic's edge was business customers. Its models, called Claude, were sold mostly by usage rather than by subscription: 80 percent of revenue in the first half of 2026 was consumption-based. The one number that tells the story is the customer count. At the end of 2023, Anthropic had about twenty customers paying it more than $100,000 a year. On September 30, 2026, it had more than 11,900.

Every Landlord at Once

Anthropic took the opposite approach to OpenAI's. Rather than one landlord, it signed up all of them.

Amazon came first. By November 2024, Amazon had invested a total of $8 billion, and AWS became Anthropic's "primary cloud and training partner," with Anthropic helping to design Amazon's Trainium chips. In October 2025, Anthropic agreed to use "up to one million" of Google's TPU chips, "well over a gigawatt of capacity online in 2026," in a deal "worth tens of billions of dollars." A month later, it "committed to purchase $30 billion of Azure compute capacity" from Microsoft, while Nvidia agreed to invest "up to $10 billion" in Anthropic and Microsoft "up to $5 billion." By 2026, it was also buying from CoreWeave and from Musk's SpaceX.AI.

The private rounds grew with the revenue. Anthropic raised $30 billion at a $380 billion valuation in February 2026, then $65 billion at $965 billion in May 2026, when its run-rate revenue passed $47 billion a year.The S-1 shows smaller amounts than the announcements: Series F-1, $13.0 billion at $140.9676 a share (August–September 2025); Series G-1, $23.45 billion at $259.1364 (January–February 2026); Series H-1, $45.73 billion at $589.0095 (May–June 2026), of which Amazon put in $5.0 billion. The announced rounds were $30 billion and $65 billion. The gap is unexplained, likely other tranches or instruments. The Series F valuation of $183 billion is from Anthropic's announcement. CNBC reported, citing sources, that run-rate revenue reached $65 billion in July 2026. "Anthropic is building the bridge between where enterprise AI stands today and where it's headed," said Sequoia's Alfred Lin.

The Filing

On June 1, 2026, Anthropic confidentially submitted a draft registration statement to the SEC. It filed five amendments over the summer and fall. On October 26, 2026, it filed its S-1 publicly, and on November 2, it filed Amendment No. 1 with results through September 30.

Here is what the filing shows.

Revenue was $41 million in 2023, $386 million in 2024 and $4.585 billion in 2025. In the first nine months of 2026, it was $33.2 billion, against $2.5 billion in the same period a year earlier. No single customer accounted for more than 10 percent of revenue, and the top five were about 25 percent. About 48 percent of revenue came through cloud marketplaces, according to CNBC's reading of the filing, meaning through the same companies that sell Anthropic its computing.

The losses need a careful reading. Anthropic's operating loss for the nine months was $1.26 billion. Its net loss was $104.0 billion. The difference was mostly a $94.3 billion non-cash charge to mark its convertible notes and SAFEs, instruments sold to early investors, up to the value implied by its soaring valuation.The original S-1 of October 26 reported first-half figures; CNBC reported first-half revenue of $16.29 billion and a net loss of $80.7 billion from that version, and Google's commitment was cited at $114.5 billion before the amendment. The 2025 net loss was $42.0 billion, and the third-quarter 2026 net loss was $23.3 billion. International Business Times put third-quarter revenue at about $17 billion. In plain English, the company's success made its old investors' claims more valuable, and accounting rules required it to book that as a loss.

Then there is the number that ties Anthropic to everything else in this chapter. Anthropic disclosed "aggregate committed spend of approximately $895.4 billion" on computing, $447.5 billion of it due before the end of 2030. That includes at least $44.9 billion to Microsoft, at least about $121.3 billion to Google, plus an estimated $125 to $150 billion more for Google's chips from 2028 to 2031, at least $110.0 billion to Amazon, and up to $120.5 billion to SpaceX.AI for Nvidia-based computing. Anthropic's commitments are other companies' backlogs. When you read about Oracle's or CoreWeave's or Google's contracted revenue, part of what you are reading is the other side of this page. MoffettNathanson estimated that Anthropic alone accounted for 40 percent of Google Cloud's backlog in mid-2026, and would supply 20 percent of Google Cloud's revenue by 2028.MoffettNathanson, "Alphabet: Are They Making the Wrong Bets?" September 16, 2026. MoffettNathanson's model works from a reported $200 billion, five-year commitment by Anthropic to Google. That figure and the S-1's commitments to Google (at least about $121.3 billion, plus an estimated $125 to $150 billion for TPUs from 2028 to 2031) cover different scopes and periods and are not reconciled here. Nor is the treatment of the Broadcom leases, about $125.2 billion of which date from a March 2026 arrangement, within the S-1's $895.4 billion of committed spend. The S-1/A shows cash and equivalents of $95.6 billion and short-term investments of $36.3 billion at September 30, 2026.

One more arrangement in the filing will look familiar to anyone who lived through 2001. Anthropic leases TPU-based computing capacity from Broadcom under five-year equipment leases, booked as finance leases, with about $161.2 billion of rent in total. To help pay that rent, Anthropic may issue up to $42.0 billion of convertible notes to Broadcom. In Bandwidth, I described how Corvis's balance sheet "allowed it to provide vendor financing to its customers," and how Broadwing used that financing to buy Corvis's equipment. Here the vendor is lending its customer the money to rent the vendor's product. The difference, again, is disclosure and demand: the filing spells it out, and the capacity is being used. The shape is the same.

Anthropic is also building for itself. Its first data center "became operational this year," the filing says, in buildings "leased and operated on our behalf," with a company called Fluidstack running them. It has signed further computing deals with Nscale, for up to about $44.6 billion through 2033, and with Lambda, for about $41.3 billion through 2034.

The governance is unusual too. A Long-Term Benefit Trust elects a majority of the board through a special class of stock. A single Class F share, held by a "Founder LLC" owned by the Amodeis and five other co-founders, carries 50.1 percent of the vote. Public shareholders will own a piece of the economics, not control. The Amodeis' pay for 2025, according to CNBC, was about $17.96 million for Dario and about $16.4 million for Daniela.

The IPO: What Is Confirmed and What Is Reported

As of today, November 6, 2026, here is what is confirmed by the SEC filings and by Anthropic itself. Anthropic has filed publicly to sell Class A common stock on the Nasdaq Global Select Market under the ticker "ANT." Twenty-two underwriters are named, led by Goldman Sachs and Morgan Stanley. The price range, the number of shares and the proceeds are all blank in the filing. The SEC has not declared the registration effective, and no final prospectus has been filed. No listing date has been set. Anthropic has not announced a postponement.

Here is what has been reported but not confirmed. The Wall Street Journal reported in October that Anthropic was targeting a raise of more than $100 billion at a valuation of $1.5 to $2 trillion, with a listing the week of November 16. Bloomberg reported a roadshow as early as the week of November 9. After the November 5 selloff, Bloomberg reported that some prospective investors were discussing a valuation of $1.2 to $1.3 trillion, down from the $1.4 to $1.5 trillion communicated to bankers and up to $2 trillion in early October. The Journal reported that three prospective investors said the price likely needs to come down, one of them to about $1 trillion, and that a smaller raise or a delay was possible. Reuters reported that investors wanted October revenue figures, which the amendment left blank, and that the roadshow could struggle to launch. Anthropic declined to comment.

This draft will be updated when the offering prices and the stock begins trading.

Step back and look at the arc. In May 2021, a group of researchers raised $124 million. In May 2026, investors valued their company at $965 billion. In October 2026, the talk was $2 trillion. On the day I write this, it is $1.2 to $1.3 trillion, and falling with the market. Exodus, the Gold Rush company of Chapter III, went from a room with no ceiling to $20 billion of Market Value. Anthropic's valuation talk fell by more than that in a single news cycle.

The bulls will point to the revenue line, which is the steepest in this book. The skeptics will point to the commitment line, $895.4 billion against $33.2 billion of nine-month revenue, and to a company whose largest suppliers are also among its largest investors and its largest sales channels. Both are reading the same filing. The market is about to tell us which one it believes.

Midnight#

Look at the week that ended today.

On Tuesday, voters rejected data centers in forty of forty-one local ballot measures, and Texas voted out the governor who had halted data-center permits only weeks before. On Wednesday, Nvidia lost its title as the world's most valuable company. On Thursday, the Nasdaq had its worst day in nineteen months, and the chip stocks fell more than 12 percent. CoreWeave was down 57.6 percent since May 8. On Friday, Bloomberg reported that Anthropic's IPO valuation talk had slipped from $2 trillion toward $1.2 trillion, in the same month it had reportedly hoped to start its roadshow. CoreWeave reports on November 10, Nvidia around November 17.

And yet. On every earnings call in this chapter, someone said demand exceeds supply. The revenue is real. The tokens are being used. The chips are running hot in Memphis, Abilene and Louisiana. The reactor at Three Mile Island is coming back.

In Bandwidth, I closed with Buffett's description of a bubble as Cinderella's ball. The guests know the carriage will turn into a pumpkin. "The giddy participants all plan to leave just seconds before midnight," Buffett said. "There's a problem, though: They are dancing in a room in which the clocks have no hands." My advice then was to leave the party before midnight. And always remember, Cash Flow is King.

Jensen Huang's version is "compute is revenue." Maybe he's right. Jim Crowe's line, which opens this chapter, says both can be true at the same time. In the short term, reality underperforms the hype. In the long term, our imaginations fail us. In 2001, the short term arrived and wiped out most of the companies that built the networks. In the long term, those networks carried the cloud, and now they carry AI.

So which half of Crowe's line are we living in? The clocks still have no hands. The music is still playing. Some of the guests are already looking for their coats.

Are we before midnight, or after?

Table 8.1: Cash Capital Spending of the Four Big Clouds, 2023–2026#

Source: SEC EDGAR XBRL data (cash purchases of property and equipment), retrieved November 6, 2026; guidance from Q3 2026 releases and calls. See note 11.

Two companies in this book were founded in Silicon Valley in the 1990s to house the Internet's computers.

The first was Exodus. It started in a Sunnyvale office whose founders knocked out the ceiling to keep forty servers cool. It went public in 1998, rose 84 percent on its first day and reached a Market Value of more than $20 billion in February 2000. It built and it bought, on borrowed money, to meet a forecast. Twenty-one months later, twenty-eight of its data centers sold for $560 million, and the company was gone.

The second was Equinix. It started with two veterans of a Palo Alto Internet exchange and a bet that sounded too modest for its time: own the room where the networks meet, and let them meet each other. It went public five months after the Nasdaq peaked and traded for 30 cents a share in 2002. It survived on a rescue from Singapore and bought finished buildings from the wreckage for a fraction of what they had cost. By 2017 it ran 175 data centers in forty-four markets.

Same industry, same decade, same Valley. One company bet on the forecast. The other bet on the room. That is the whole book in two sentences, and it is where the scoring has to start.

Keeping Score#

In Bandwidth's last chapters, I scored the four national fiber providers that survived the Reconsolidation, by their Market Value, their debt and the structural problems they carried. Let's do the same here, with one adjustment. The data-center business has two sets of players to score. One set has been through a Bust. The other hasn't yet.

The Survivors

Equinix is the meeting room. It owns the cross-connects, the smallest product in the business and the hardest to replace, because a network that leaves an Equinix building leaves behind every other network it connected to there. It grew through the 2008 crisis, booking $704.7 million of revenue that year, up 68 percent, while banks failed around it. In the Reconsolidation it bought Telecity and twenty-nine of Verizon's data centers. When the cloud companies began asking for whole buildings, it served them through xScale, using mostly a sovereign fund's money rather than its own balance sheet. In early November 2026, Equinix's Market Value was about $100 billion.About $100–104 billion, depending on the day's close and share count (Benzinga and Macrotrends, November 3–4, 2026). Re-check at publication. From 30 cents a share to $100 billion. Not bad for a meeting room.

In late October 2026, as the AI stocks around it fell, MoffettNathanson's Del Deo called Equinix's interconnection model "durable through business and technology cycles." "Equinix is a veteran," he wrote, "in a world where upstarts have, for better or worse, grabbed an outsized share of attention and growth."

Its exposure to the AI Gold Rush is real but indirect. The giant AI campuses are wholesale buildings, and when it launched xScale, Equinix said it wasn't entering the wholesale business broadly. That choice may cost it some of the Boom. It also leaves it less exposed to the Bust.

Digital Realty is the landlord. GI Partners turned $280 million of equity and twenty-five unwanted buildings into the first data-center REIT, worth about $6 billion by 2009, when its 11 million square feet were 95 percent leased. Between 2015 and 2020 it bought Telx, DuPont Fabros and Interxion, about $1.9 billion in cash and about $16 billion more in stock deals including debt. Paying in stock meant its shareholders shared the risk with the sellers. At September 30, 2026, Digital Realty carried about $19.5 billion of debt, at 5.2 times its adjusted EBITDA.Digital Realty, "Digital Realty Reports Third Quarter 2026 Results," press release, October 29, 2026: about $19.5 billion of total debt, of which $17.9 billion was unsecured, and net debt at 5.2 times adjusted EBITDA. Re-check at publication. Its returns tell the story of two eras. In the cloud years, its return on capital fell more than 3 percentage points from 2018 to 2023. In October 2026, Del Deo wrote that AI demand and pricing had "flipped that script" for the landlord.

Digital Realty's question is the one Chapter VI raised: what happens when your tenants know your costs to the dollar and can build without you? In the AI Gold Rush, the answer so far has been good news for any landlord with powered land. If the Boom slows, the same tenants who signed the leases will be negotiating the renewals.

The cloud companies are the strongest and the most changed. For most of this book they were the best cash machines in business history. In 2026, for the first time, their spending ran past their cash. Amazon spent $77.3 billion on property and equipment in the third quarter of 2026 alone, and its trailing free cash flow through September was an outflow of $42.4 billion. Alphabet's free cash flow was negative in the second and third quarters. Meta financed most of its Louisiana campus through a vehicle that kept about $27 billion of debt off its balance sheet. Microsoft guided to about $170 billion of spending for calendar 2027.

And yet the revenue is real. Azure grew 48 percent in a quarter. AWS grew 39 percent. These are still the best cash machines in business history, and they can afford a bad year, or even a bad few years. The question for them is whether the buildings they are paying for today will earn their cost before the chips inside them are obsolete.

Private infrastructure capital bought the industry at the top of the last cycle's prices. Four companies, QTS, CyrusOne, CoreSite and Switch, changed hands for about $46 billion in eleven months. CoreSite went for about thirty times EBITDA, priced like a tower. The buyers' thesis was a stream of rent from the best credit in the world. That thesis holds as long as the tenants are Amazon, Microsoft, Alphabet and Meta. It gets harder to defend when the tenant behind the tenant is a model lab raising money every few months.

And the private owners have started looking for the exit. In October 2026, Del Deo reported that CyrusOne and Switch, both taken private in the Reconsolidation, were considering a return to the stock market, along with Vantage, at the top of the AI cycle. The window was already shaky. Del Deo noted "a little wobbliness" in the IPO market for AI-linked infrastructure, and a Blackstone data-center vehicle that had listed in May 2026 traded below the cash it had raised. He titled that note "90 Cents on the Dollar."

The Newcomers

Nvidia is the shovel seller, and the shovel seller's profits are the one thing in Chapter VIII nobody disputes. Its data-center revenue went from about $15.0 billion in the fiscal year ended January 2023 to $193.7 billion three years later, and its gross margin in the July 2026 quarter was 75.0 percent. It returned $41.1 billion to its shareholders in a single year. As of November 6, 2026, it was worth about $4.5 trillion. Its weakness is the one Huang himself let slip: concentration. A large share of its demand has come from a handful of buyers, and Nvidia has invested in several of them, including $30 billion in OpenAI's 2026 round, up to $10 billion in Anthropic and a reported anchor order in CoreWeave's IPO.

Oracle is the builder with the backlog. Its remaining performance obligations reached $664.0 billion by August 2026, and its debt rose from $92.9 billion to $130.1 billion in a year. About $300 billion of the backlog is a single contract with OpenAI. The price of insuring its bonds against default climbed to the highest level since 2009. Oracle has a real software business underneath, which the fiber companies of 2000 did not. But it is borrowing to serve a customer whose own money comes from investors.

CoreWeave is this cycle's Exodus, or its Global Crossing, or neither. It guided to $12.4 to $13.2 billion of revenue for 2026 and $35 to $39 billion of capital spending, about three times its revenue. Its debt was $35 billion by CNBC's count and $51.6 billion by Trefis's. Its trailing free cash flow was negative $13.66 billion. Its bonds yielded above 15 percent, and from May 8 to November 4, 2026, its stock fell 57.6 percent. Its first big customers were among the most creditworthy companies on earth, but much of its newer backlog came from customers without investment-grade credit, OpenAI first among them. Its collateral is chips, and nobody yet knows what a three-year-old GPU is worth in a downturn.

Two weeks before that low, Del Deo published the sharpest analyst verdict of this cycle, under a title from Unforgiven: "Deserve's Got Nothing to Do with It." CoreWeave, he wrote, was not "circling the drain." Its customers were short of computing, and its projects made money. But its obligations were "staggering," and it lacked the cash flow and credit ratings of the cloud giants it wanted to join. "Capital markets can be mercurial, unsentimental, quick to judge, and violent. That's the risk CoreWeave faces." His one consolation for the bulls was what he called the "Nvidia Put": if CoreWeave got in trouble, "we could very well see the 'Nvidia Put' be exercised."

The model labs are the customers whose commitments fill everyone else's order books. OpenAI raised $122 billion at an $852 billion valuation in March 2026, and every large check came from a company that also sold it chips or cloud capacity. xAI was folded into SpaceX in a deal Bloomberg valued at $1.25 trillion. Anthropic booked $33.2 billion of revenue in the first nine months of 2026, against $2.5 billion a year earlier, the steepest line in this book. It also disclosed about $895.4 billion of committed spending on computing, $447.5 billion of it due before the end of 2030. As this draft was finished, its IPO valuation talk had slipped from as much as $2 trillion toward $1.2 to $1.3 trillion.

Table 9.1: The Scorecard

Read the right-hand column again. The survivors' strength is measured in buildings and tenants. The newcomers' strength is measured in revenue growth, backlog and commitments, which is to say in promises, some of them very good promises. In Bandwidth, the companies that failed were the ones that confused building with earning. The scoreboard can't yet tell us which of the newcomers are doing which.

How It Might End#

I won't predict the hour. But I can lay out the ways this cycle could play out, and how likely each looks from what's in this book.

Most likely: a shakeout, short of a collapse. The demand is real, and it keeps growing. But it grows more slowly than the commitments assumed, and the companies at the end of the longest chains of borrowed money get squeezed first. A lab misses a funding round, a GPU renter can't refinance its 15 percent debt, a backlog gets renegotiated. Some of today's Gold Rush companies don't make it, and their assets change hands. The cloud companies slow their spending without stopping it. The buildings stay full, though some of the chips in them are worth far less than their owners paid. This is Chapter IV without the 22 percent occupancy. The Bust would hit the balance sheets and spare the buildings.

Possible but not likely: a repeat of 2001. Everything falls at once. The labs can't raise money, their commitments to the clouds and to Oracle and CoreWeave go unpaid, Nvidia's orders collapse, and the half-built gigawatt campuses sit as empty as the telecom hotels of 2001. What makes this unlikely is the difference Chapter VIII kept returning to: the revenue is real, the customers are using the computing, and the biggest buyers have hundreds of billions of dollars of cash flow from other businesses. What keeps it possible is the circle. When suppliers finance their customers and the customers' commitments become the suppliers' backlogs, a crack in one place can travel quickly.

Plausible: power and politics apply the brakes. The scarce thing in this cycle is electricity, and it can't be conjured on a press release's schedule. PJM came up short of reliable capacity three auctions in a row. Grid connections take five years. Texas paused new connections and permits before its voters changed governors. On November 3, 2026, voters rejected data centers in forty of forty-one local ballot measures. Virginia's own auditors recommended that utilities "delay, but not deny" service to big new loads, and Ohio made data centers guarantee most of what they reserve. The forecasts themselves deserve humility in both directions. In 2000, a witness told Congress that Mark Mills's Internet power forecast was "wrong by at least a factor of 8," and the critics were right. In 2016, Lawrence Berkeley forecast little growth, and AI broke the trend. In 2025, Mills was back, and a consumer advocate told the same committee that forecasters "consistently overestimate." In this scenario, the Boom ends with a long, slow squeeze rather than a pop. That would be painful for the developers waiting on power. It would also prevent the kind of overbuilding that sank the last cycle, and it would make every building that already has power worth more. The irony would be hard to miss: the neighbors who voted no might end up protecting the industry from itself.

Unlikely but plausible: the forecasts turn out too small. Jim Crowe's line opened Chapter VIII: "In the short term, reality underperforms the hype. In the long term, our imaginations fail us." In 2001, the short term arrived first and wiped out most of the companies that built the networks. Then the long term arrived, and the networks carried the cloud. It is possible that this time the long term arrives early, that "compute is revenue" holds, and that a cash-flow valuer like Aswath Damodaran turns out to have been conservative. Nobody should rule it out. Nobody should bet the balance sheet on it, either.

And after any of them, a Phoenix. If a Bust comes, the pattern says who wins next. Buyers with cash and patience will walk through the wreckage, and they'll be shopping for what lasts. Del Deo's comparison of the two booms says where to look. The fiber and conduit of 2000 had decades of life, and much of it, he wrote, "is supporting today's AI Boom!" This time most of the money went into chips that last years. A data center lasts for decades. A powered site with a grid connection may be the scarcest asset of all. A three-year-old GPU may be worth whatever the market says that week. The likeliest buyers are the ones already standing: Equinix and Digital Realty, the infrastructure funds with committed capital and the cloud companies themselves. In 2004, Savvis paid $155 million for fifteen of Exodus's data centers. Somebody will get the 2026 version of that deal.

The Lessons of Two Cycles#

I've now watched the same cycle play out in fiber and in data centers, and read about it in time-sharing. A few lessons hold every time.

The building survives. The balance sheet might not. Exodus went bankrupt; its buildings kept humming, and the customers stayed. Cable & Wireless USA went bankrupt; the customers stayed again. In 2008, DuPont Fabros ran out of lenders and kept its tenants. When you value a data-center company, separate the building from the borrowing.

Know what is backing the debt. In the hosting Gold Rush, the forecast was the asset, and when it came down, so did the companies. In the Phoenix, the lease was the asset, and it held through the 2008 crisis. In the AI Gold Rush, the asset is the contract. A signed contract is far better than a forecast. It is also only as good as the customer's ability to pay, and some of the largest customers in this book pay with money they raised a few months earlier.

Own the place where everyone meets. MAE-East began as a free handshake in a Virginia office building. The early operators of the exchanges gave away the most valuable part. Equinix figured out that a meeting place gets more valuable with every tenant who moves in, and that no competitor with a bigger budget can copy it. At MFS, we thought we were selling roads. The intersection turned out to be the prize.

Watch the customer who can build without you. In 1968, IBM's customers worried about competing with their own landlord. By 2016, the cloud companies had learned to build their own buildings, and the power shifted from the people who own data centers to the people who fill them. In the AI era, the companies selling the chips also finance the buyers. Ask who depends on whom, and what happens to each if the other changes its mind.

Don't build on borrowed money when the cost of money turns. In 1969, University Computing blamed its earnings miss on "curtailed expansion of the company's computer network facilities, due to high interest rates." In 2008, DuPont Fabros said the banks it was talking to had "merged or disappeared." CoreWeave paid 10 to 17 percent for the loans that bought its chips. The sentence keeps coming back because the mistake keeps coming back.

Be wary when the seller finances the buyer. Global Crossing's suppliers and customers were often the same companies. In this cycle, Nvidia invested in CoreWeave, sold it every GPU it ran and bought computing from it, and Anthropic's IPO filing shows it may issue up to $42.0 billion of convertible notes to Broadcom to help pay Broadcom's lease rent. The circle is bigger, better disclosed and attached to real computing, and nothing in this book's sources suggests anything like the frauds of 2001. But revenue that depends on your customer's next funding round deserves a second look.

Mind the meter. In 2001, one hundred server farms asked PG&E for 1,200 megawatts, and the utility told Qwest to wait until 2002. The warning faded when the demand did. Twenty-five years later, single campuses ask for gigawatts, PJM's capacity price has risen more than elevenfold in four auctions, and Texas has written a kill switch into law. A building without power is an empty shell. A building with power, in this cycle, is a gold mine.

The neighbors get a vote. For most of this book, data centers were the buildings nobody saw. By its Chamber of Commerce's count, Loudoun County collected nearly $1.2 billion of revenue from them and cut its property tax rate. Then the bills, the turbines and the water became visible, and the neighbors noticed. Every Boom needs the community's consent. This one learned that lesson in a single election.

Patience pays, and it pays the patient. Keydata took four and a half years to post its first profitable month and outlasted most of the time-sharing Boom. Rackspace went public eight years late, and profitable. GI Partners turned $280 million into more than $1 billion by buying what nobody wanted. In every cycle in this book, the biggest fortunes went to the people who bought after the Bust. The people who built before it mostly paid for the party.

Before Midnight#

In Bandwidth, I closed with Buffett's fairy tale. A Boom, he said, is "akin to Cinderella's Ball," an elaborate party whose guests stay late because "they hate to miss a single minute of what is one helluva party." Every cycle in this book had its ball.

In 1968, if you mentioned time-sharing in a Wall Street bar, the bankers would buy you a drink and offer you a million dollars. In the summer of 2000, a telecom-hotel developer explained that eighteen months "is like 10 years in Internet time," and Morgan Stanley put $1.25 billion behind the telecom hotels. In January 2025, Masayoshi Son stood at the White House and called Stargate "the beginning of our golden age." For a few hours in September 2025, Larry Ellison was the richest man in the world.

Cinderella knew her carriage would turn into a pumpkin, and she knew when. She could watch the hands of the clock. The guests at Buffett's ball can't. "The giddy participants all plan to leave just seconds before midnight," he said. "There's a problem, though: They are dancing in a room in which the clocks have no hands."

The time-sharing dancers of 1969 didn't see midnight coming. Neither did the hosting crowd of 2000, or the fiber crowd I danced with. As I write this, in the first week of November 2026, the music is still playing. The chips are running hot in Memphis, Abilene and Louisiana. The reactor at Three Mile Island is coming back. And some of the guests are already looking for their coats.

I don't know what time it is. Nobody does. That is Buffett's point. But I know what the clocks looked like the last two times, and I know who was still standing when the music stopped. They were the ones who owned the room where everyone met, who signed real tenants to real leases, who didn't borrow against a forecast and who had cash when everyone else needed it.

So heed the words of the Oracle from Omaha. Be greedy when others are fearful. Be fearful when others are greedy. Know there are no hands on the clock. Leave the party before midnight.

Own the room where everyone meets. Never mistake a forecast for a lease. Ask who is paying the payer. Secure the power before you pour the concrete. Buy the building after the Bust.

And always remember, Cash Flow is King. This time around, so is Power.

Dates are those reported in the chapters. Where a deal was announced in one month and closed in another, the entry gives the date that matters to the story and the chapter gives the rest.

The Prologue Cycle, 1962–1990#

The First Cycle, 1987–2022#

The Second Cycle, 2022–November 2026#

Prices are as reported in the chapters. Where the announced and closing values differ, both are given. "Including debt" means the price was reported as an enterprise value. Deal-structure detail and source conflicts are in the chapter footnotes.

The Prologue Cycle#

The Networks Meet#

The Hosting Gold Rush#

Empty Floors#

The Phoenix#

The Cloud#

The Reconsolidation#

The AI Gold Rush#

What the Prices Said#

The prologue. The first prices in this book were paid in paper. University Computing used a stock valued at more than 100 times earnings to buy an insurance company, and Perot offered EDS stock for a radio maker with 55 times his sales. When GM paid $2.55 billion for EDS in 1984, it paid mostly in cash for a business with real contracts, and the stock market of 1970 had priced in most of that value 14 years early.

The networks meet. The early prices were small and then suddenly very large. The NSF obligated $211,587 for the Washington NAP. Crowe offered $8 million for UUNET in 1993 and paid $2 billion three years later, and WorldCom paid $14.3 billion for MFS a few months after that. The buyers were paying for the meeting places and the traffic, though most of them still thought they were buying networks.

The Gold Rush. The prices of 1999 and 2000 were set by forecasts and paid in stock. When news of the GlobalCenter talks leaked, Exodus's shares jumped 49 percent, and because Exodus was paying in shares, the price of the deal jumped with them. Digex, worth $5.3 billion on the market, drew an $8.5 billion offer. The currency was as inflated as the target.

The Bust. The Bust repriced everything at once. Exodus's best data centers sold for less than 3 percent of its peak Market Value, Data Return for about one-fifteenth of its IPO valuation and PSINet's U.S. business for about $10 million against nearly $3 billion of debt. The buildings changed hands at a fraction of their cost, and the customers stayed in them.

The Phoenix. The Phoenix prices were the cheapest in the book and the best trades. Equinix paid about $900,000 for the guts of a finished building that could bring in $30 million a year. GI Partners turned $280 million of equity into more than $1 billion. By 2011, the same buildings were being sold to the phone companies at prices that looked a lot like another top.

The cloud. The cloud era added a new kind of price: the IPO of a wholesale landlord, valued like real estate. Digital Realty, DuPont Fabros, CoreSite, CyrusOne and QTS all went public between $12 and $21 a share. Zayo's own zColo purchases, from MarquisNet to Latisys, show what a fiber company paid to own the rooms at the ends of its fiber.

The Reconsolidation. By 2021, the buyers were paying for cash flow, and they were paying with cheap capital. QTS sold for about 3.7 times its IPO price, CyrusOne for almost five times and CoreSite for more than ten times. American Tower paid about 30 times CoreSite's EBITDA, a tower multiple for a data center. Four companies changed hands for about $46 billion in 11 months.

The AI Gold Rush. The prices of this cycle are mostly investments, and many of the investors also sell to the companies they back. Nvidia invested in OpenAI, Anthropic and CoreWeave's IPO. Amazon, Microsoft and Google invested in Anthropic while selling it computing. Anthropic went from a $124 million Series A in 2021 to a $965 billion valuation in May 2026, and in the first week of November 2026 the market began to argue about what it was worth.

In Bandwidth*, I tallied the fiber companies by what they were worth at the top and what was left at the bottom. This is the same roll-call for the companies in this book. Every figure comes from the chapters, with the date it was reported. Where the record shows a rise from a low point rather than a fall from a high one, the row reads in that direction, and the note says so. Market Values for the AI-era companies move daily; these are the figures as of the first week of November 2026.*

The First Bust#

The Hosting Bust#

The Phoenix and the Cloud#

The Reconsolidation#

The AI Gold Rush, So Far#

Not in the Roll-Call#

Several companies in this book appear without both ends of the pair in the sources. Akamai was worth more than $13 billion at the close of its first day of trading (October 29, 1999), but the chapters record no trough. Level 3 passed $50 billion, and Bandwidth tells the rest. The trough for University Computing's stock is also missing: the last quote in the sources is 24 3/8, on May 1, 1972.

In 1983, Warren Buffett drafted a set of business principles, and in 1996 he published them as an Owner's Manual in Berkshire Hathaway's annual report. In January 2014, with respect and deference to Buffett, I published an Owner's Manual for Zayo's investors. Bandwidth reprinted it as its first appendix.

This one is shorter, and it is written for a different reader: anyone who owns, lends to, builds or rents a data center as the second cycle in this book plays out. The principles come from the stories in the chapters. Most of them have played out more than once.

Principle 1: Cash Flow is King#

Buffett defines the value of a business as the cash it will produce "between now and doomsday." Revenue growth, backlog, square feet and megawatts are all ways of guessing at that number, and none of them is the number. In the year Lehman Brothers failed, Equinix grew its revenue 68 percent and earned a profit, while DuPont Fabros, still building on borrowed money, halted construction.

Principle 2: Never mistake a forecast for a lease#

A forecast is somebody's opinion of demand. A lease is a customer's promise to pay. Exodus went public on Forrester's forecast that hosting would grow eighty-fold in five years; Digital Realty went public six years later, at the same $12 a share as Equinix, on buildings that were 87.1 percent leased.

Principle 3: Separate the building from the borrowing#

A data center can be full and still sink the company that owns it. Value the building and the balance sheet separately, and ask which one the debt depends on. Globix said its centers were 95 percent full in June 2001, and seven months later it was headed for bankruptcy with a Market Value under $6 million.

Principle 4: Own the place where everyone meets#

A network can be duplicated by anyone with a bigger budget. A meeting place gets more valuable with every tenant who moves in, and a competitor can't copy the tenants. At MFS we thought we were selling roads, and the intersection we built at MAE-East turned out to be the prize; Equinix built a company worth about $100 billion on the same idea.

Principle 5: Know which customers can build without you#

A landlord whose tenants can build their own buildings has to earn the renewal every time. When the biggest customers know your costs to the dollar, your pricing power lasts only as long as they can't build fast enough. Facebook published the designs for its Prineville building, and one of its engineers explained why giving them away cost nothing: "We don't hold this technology as a key to our business."

Principle 6: Don't build on borrowed money when the cost of money turns#

Capital-hungry builders are always one rate increase or one frozen credit market away from stopping. In 1969, University Computing blamed its missed forecast on "curtailed expansion of the company's computer network facilities, due to high interest rates"; in 2008, DuPont Fabros halted construction because the banks it had been talking to "merged or disappeared."

Principle 7: Ask who is paying the payer#

A contract is only as good as the customer's ability to pay, and some customers pay with money they raised a few months earlier. When suppliers invest in their own customers, follow the money around the circle before you count the revenue. Anthropic disclosed about $895.4 billion of committed spending on computing, and part of every backlog reported by its suppliers is the other side of that page.

Principle 8: Secure the power before you pour the concrete#

A building without power is an empty shell, and power arrives on the utility's schedule. In 2001, PG&E told Qwest it couldn't power a Sunnyvale data center until 2002; by 2026, PJM had come up short of reliable capacity in three auctions in a row, and the median power project took 61 months to connect to the grid.

Principle 9: The neighbors get a vote#

A data center needs its community's consent, even when it never asks. Pay your own way on power and water, and say so in public before someone else says otherwise. On November 3, 2026, voters rejected data centers in 40 of 41 local ballot measures, and even Loudoun County, which collected nearly $1.2 billion of revenue from them, had already ended by-right approvals.

Principle 10: Measure the thing itself#

Every Boom has a number that everyone quotes and nobody checks. Find the person measuring actual use, and listen when the measurement disagrees with the forecast. In 1998, Andrew Odlyzko showed that Internet traffic was doubling about once a year, not every hundred days, and his inconvenient truth was ignored by the people raising money on the myth.

Principle 11: Buy the building after the Bust#

The best trades in this book were made by buyers with cash and patience who walked through the wreckage. Pay for what lasts, and let someone else pay for the party. Equinix bought the inside of a finished San Jose data center for about $900,000, and GI Partners turned $280 million of equity in unwanted buildings into more than $1 billion.

Principle 12: Be candid, especially when the news is bad#

Investors, lenders and neighbors are owners too, and they deserve the same information management uses to run the business. Triumphs get trumpeted; mistakes deserve the same airtime. At the top of the hosting Boom, Level 3's Jimmy Byrd warned in the Wall Street Journal that too many data centers were coming online at once, and the record shows he was right.

Plain-English definitions of the technical and financial terms used in this book. Where a term has a defining example in the chapters, the example is given.

• AI factory. Jensen Huang's name for a data center built to train and run artificial-intelligence models, one giant computer to a building, whose product is tokens.
• Backlog. Revenue a company has signed contracts for but not yet delivered. Oracle and CoreWeave report it as remaining performance obligations. A backlog is a promise to pay, and it is only as good as the customer behind it.
• Behind-the-meter. Taking electricity directly from a power plant next door without passing it through the public grid. In 2024, FERC rejected an expanded behind-the-meter deal between Amazon and Talen's Susquehanna nuclear plant, over the question of who pays for the grid.
• Capacity auction. An annual auction in which a grid operator pays power plants to be available a few years in the future. PJM, the operator for 13 states and Washington, D.C., runs the best-known one. Its price, in dollars per megawatt-day, shows how tight the grid is.
• Capital spending. Money spent on long-lived assets such as buildings, power equipment and chips, also called capex. In the filings quoted in this book, it appears as purchases of property and equipment.
• Carrier hotel. A building, often an old telegraph, telephone or industrial building, where many network carriers rent space for their equipment and connect to one another. Also called a telecom hotel or telco hotel. 60 Hudson Street, 111 Eighth Avenue and One Wilshire are the classic examples.
• Chapter 11. The section of the U.S. Bankruptcy Code under which a company keeps operating while it reorganizes. Its creditors often end up owning the company that emerges. Under Chapter 7, by contrast, a company is usually liquidated.
• Cloud. Computing rented over a network, by the hour or by use, from a provider that owns or leases the data centers. Infrastructure as a service (IaaS) is the most basic form: renting raw computers and storage, as Amazon began doing with S3 and EC2 in 2006.
• Colocation. Renting space, power and cooling inside someone else's data center for your own equipment. Often shortened to colo, a nickname the Wall Street Journal explained to its readers in 1998.
• Credit default swap. A contract that pays out if a borrower defaults. Its price works as the market's measure of how risky a company's bonds are.
• Cross-connect. A physical cable between two tenants' equipment inside the same data center. It is the smallest product in the business and one of the most profitable.
• Data center. A building, or part of one, designed to house computers, with reliable power, backup generators and batteries, cooling, physical security and connections to outside networks.
• Depreciation. The accounting charge that spreads the cost of an asset over its useful life. Assume a longer life and the annual charge shrinks, which raises reported earnings. Whether a GPU lasts six years or two is one of the central arguments of the AI Boom.
• DIP financing. Debtor-in-possession financing: a loan made to a company while it is in Chapter 11, usually repaid ahead of the company's older debts.
• EBITDA. Earnings before interest, taxes, depreciation and amortization. Bandwidth called it Profitability. It measures the cash a business generates before paying for capital spending, interest and taxes.
• Enterprise Value. Market Value plus net debt: the price of the whole company, owed to shareholders and lenders together. Deal prices reported "including debt" are enterprise values.
• Facilities management. Running a customer's computer operations for it under a long contract, in the customer's building or the operator's. EDS and IBM built big businesses this way. Today it is usually called outsourcing.
• FFO and AFFO. Funds from operations and adjusted funds from operations, the profit measures REITs report in place of earnings. FFO adds depreciation back to net income. AFFO also subtracts the recurring capital spending needed to keep the buildings running.
• Gigawatt. A thousand megawatts, or a billion watts. A bit more than the output of the restarted reactor at Three Mile Island. In the AI era, single campuses are measured in gigawatts.
• Glass house. The industry's later name for the centralized, company-owned computer room of the mainframe era, with its raised floor, air-conditioning and guards.
• GPU. Graphics processing unit, a chip that does thousands of small calculations at once. Built for video games, it turned out to suit the training of neural networks, and Nvidia's GPUs became the scarcest commodity of the AI Boom.
• Hot site. A ready-to-use computer room kept empty for a customer whose own data center is destroyed. Comdisco and SunGard led the business in the 1980s. An empty shell, with space and no equipment, cost far less.
• Hyperscale. The largest buildings and campuses, built for or by the handful of cloud and AI companies and measured in tens or hundreds of megawatts, and now in gigawatts. The companies themselves are called hyperscalers.
• Interconnection. Connecting networks and customers to one another inside a building, through a shared switch or a cross-connect. Equinix built its business on it.
• Interconnection queue. The waiting line of power projects asking to connect to the grid. At the end of 2025, about 2,061 gigawatts were in the U.S. queue, and the median project took 61 months to get through it.
• Internet exchange point (IXP). A place where networks swap traffic, usually across a shared switch. MAE-East, MAE-West and PAIX were IXPs.
• MAE. Metropolitan Area Exchange, MFS's name for its Internet exchanges. MAE-East in Tysons Corner, Virginia, set up in 1992, and MAE-West in San Jose anchored the two coasts.
• Managed hosting. Running a customer's servers for it inside the host's data center, including the power, cooling, network and staff. Exodus, Savvis and Rackspace were managed-hosting companies.
• Market Value. The value of all of a company's shares, also called market capitalization or equity value. This book capitalizes it as a defined term, as Bandwidth did.
• Meet-me room. The room inside a carrier hotel where networks physically connect to one another. Zayo's zColo was the exclusive operator of the one at 60 Hudson Street.
• Megawatt. A million watts. Data centers are sized by the power they can deliver to their computers, and wholesale tenants often lease a megawatt or more. A kilowatt per rack, or per cabinet, measures how densely the computers are packed; the SuperNAP offered up to 17.2 kilowatts per cabinet in 2010.
• Network Access Point (NAP). The NSF's name for the four exchanges it designated in 1994 to carry traffic between commercial networks after NSFNET shut down. MFS Datanet ran the one in Washington.
• Off-balance-sheet financing. Borrowing through a separate vehicle so that the debt doesn't appear on the sponsoring company's balance sheet. Meta financed much of its Hyperion campus this way.
• Peering. Two networks agreeing to exchange traffic, originally without charging each other. UUNET began charging some of its peers in 1997.
• Perpetual capital. Money that doesn't have to be returned to investors on a fixed schedule, so it can hold an asset indefinitely. Blackstone used it to buy QTS.
• Powered shell. A building with its structure, utility power and connections in place, but without the cooling, generators and fit-out a tenant adds itself. It sits at the cheap end of wholesale.
• Power usage effectiveness (PUE). The total power a data center draws divided by the power its computers use. A PUE of 1.5 means half a watt of overhead for every watt of computing. Facebook's Prineville building reported 1.07.
• Public benefit corporation. A company whose charter names a public benefit that its directors are legally required to weigh alongside profit. Anthropic is one.
• Raised floor. The elevated floor of a computer room, with cables and cooled air running underneath. The phrase came to stand for data-center space itself.
• REIT. Real estate investment trust: a company that owns property and passes most of its rental income through to shareholders, in return for paying little corporate income tax. Digital Realty became the first data-center REIT in 2004.
• Retail colocation. Renting a cage or a few racks, buying power by the circuit and paying for cross-connects to other tenants. Equinix is the archetype.
• Round-tripping. Two companies selling each other capacity mainly to inflate both sides' revenue. Bandwidth told how Global Crossing and Qwest did it, and at Qwest it was part of a fraud.
• Sale-leaseback. Selling a property and immediately leasing it back, to raise cash while keeping the use of the building. Savvis sold several former Exodus data centers to Du Pont Fabros Interests this way in 2004.
• Section 363 sale. A sale of a bankrupt company's assets, usually by auction, under that section of the Bankruptcy Code. Savvis won the remains of Exodus in one.
• Small modular reactor (SMR). A nuclear reactor smaller than a conventional plant, designed to be built in a factory and installed in units. Google and Amazon signed deals for them in 2024, none yet built.
• Take-private. The purchase of a public company by private investors, which removes its shares from the stock market. Between 2021 and 2022, QTS, CyrusOne and Switch were taken private.
• Time-sharing. Dividing one large computer among many users at once, each typing at a terminal over a phone line and paying by the hour. The business model of Tymshare and GE in the 1960s, and of the cloud 40 years later.
• Tokens. The small chunks of text an AI model reads and writes. AI services are often priced by the token.
• Training and inference. The two jobs of an AI data center. Training builds a model by running enormous amounts of data through it; inference runs the finished model to answer users' requests.
• Wholesale colocation. Leasing a large block of space with its own dedicated power, often a megawatt or more, to one tenant for several years. DuPont Fabros was the archetype.

Grouped by the era and company in which each person first takes the stage. Titles are the ones they held in this story, and the years are the years in which they appear in it, not necessarily their full tenure.

The Glass House#

IBM

• Thomas J. Watson Jr., chairman (1964)
• Fred Brooks, the IBM executive who described the System/360's storage (1964)
• Walt Grom, IBM executive who said IBM was "happy to do the plumbing" for web hosting (1999)

Electronic Data Systems

• Ross Perot, founder (1962–1984)

University Computing Company

• Sam Wyly, chairman (1968–1972)
• Charles J. Wyly Jr., president, and Sam Wyly's brother (1969)

Time-Sharing and the Computer Utility

• Robert A. Leonard, ITT executive who forecast the "data-center industry" (1965)
• Tom O'Rourke, president of Tymshare (1968)
• John T. Gilmore and Charles W. Adams, founders of Keydata (1965–1970)
• James H. Binger, chairman of Honeywell, which bought GE's computer business (1970)
• Jeff Wilkins, founder of CompuServe (1969)
• Bill Loudin, manager of GE's GEnie service (1985)

Disaster Recovery

• Kenneth N. Pontikes, head of Comdisco (1982–1987)
• Raymond R. Hipp, president of Comdisco (1983)
• Martin A. Weinstein, head of Comdisco Equities (1987)
• Paul Catalano, founder of DataPort, beneath the World Trade Center (1986)
• Richard Reese, of Iron Mountain (1986)

The Customers

• Arthur Collins, founder of Collins Radio, who fended off Perot (1969)
• Katherine Hudson, Kodak's director of corporate information systems (1989)
• George DiNardo, of Mellon Bank, who negotiated with IBM by waving a sword (1989)

Where the Networks Meet#

The National Science Foundation

• Stephen Wolff, NSFNET program director (1994)
• Don Mitchell, NSF official who oversaw the late handoff (1994)

MFS

• Jim Crowe, chairman and chief executive, later founder of Level 3 (1993–1996)
• Royce Holland, MFS executive who assigned Dan Caruso to open local competition (1993)
• Kevin O'Hara, who hired Dan Caruso at MFS and later worked with him at Level 3 (1991–2000)
• Scott Yeager, city manager of MFS Houston, whose idea helped spark MAE-East, later a leader of Enron Broadband (1989–1992)
• Steve Feldman, MFS Datanet's NAP principal investigator and MAE's principal engineer (1994)

UUNET

• Rick Adams, founder (1987–1996)
• John Sidgmore, chief executive, later vice chairman of WorldCom (1993–2002)

PSINet

• Bill Schrader, founder and chief executive (1995–2001)

The Palo Alto Internet Exchange

• Brian Reid, Digital Equipment researcher behind the exchange (1994–1997)
• Paul Vixie, cofounder and keeper of Root Server F, later chief technology officer of Metromedia Fiber Network (1997–2001)

The Carrier Hotels

• Ken Carmel, general partner of the owner of 60 Hudson Street (1999)
• Charles Bendit and Paul Pariser, founders of Taconic Investment Partners, which bought 111 Eighth Avenue (1997–1998)

Northern Virginia

• Robert Pittman, of AOL (1996)
• David McWatters, Loudoun County supervisor (1997)
• Stephen Belomy, executive vice president of AboveNet (1999)

The Internet's Elders

• John Curran, BBN technology chief (1996)
• Bill Woodcock, of Packet Clearing House (2021)
• Vint Cerf, of MCI (1997)

The Hosting Gold Rush and the Bust#

Exodus Communications

• K.B. Chandrasekhar, cofounder and later chairman (1992–2000)
• B.V. Jagadeesh, cofounder (1992–1998)
• Ellen Hancock, president from 1998, later chairman and chief executive (1998–2001)

Equinix

• Jay Adelson, cofounder and chief technology officer (1998–2000)
• Al Avery, cofounder, later president and chief operating officer (1998–2000)
• Peter Van Camp, chief executive, later executive chairman (2000–2007)

Global Crossing

• Gary Winnick, founder (1999–2002)
• Leo Hindery, chairman and chief executive of GlobalCenter (1999–2000)

WorldCom and Digex

• Bernie Ebbers, chief executive of WorldCom (1996–2002)
• William B. Chandler III, Delaware chancellor in the Digex lawsuit (2000–2001)

The Carriers

• Joe Nacchio, chief executive of Qwest (1997–2002)
• Lew Wilks, Qwest executive (2000)
• Phil Anschutz, who replaced Nacchio (2002)
• Jimmy Byrd, Level 3 senior vice president who warned of a data-center glut (2000)
• Ron Vidal, Level 3 senior vice president (2001)
• Paul Gudonis, chief executive of Genuity (2000–2002)

Metromedia Fiber Network and AboveNet

• Steve Garofalo, founder and chief executive of Metromedia Fiber Network (1999–2002)
• Howard Finkelstein, of Metromedia Fiber Network (1999)
• Bill LaPerch, chief executive of the reorganized AboveNet (2003–2012)

The Developers

• Andrew Penson and Mark Teitelbaum, of Argent (2000)
• Manny Medina, founder of Terremark and its NAP of the Americas, later chief executive of the company that bought CenturyLink's data centers (2000–2016)

Akamai

• Tom Leighton, cofounder (1998)
• Danny Lewin, cofounder (1998–2001)
• Jonathan Seelig, cofounder (1998)
• George Conrades, chief executive (1999)
• Paul Sagan, president and chief operating officer (1999)

The Analysts and Buyers of the Bust

• Stephen Murphy, CIBC analyst who counted the empty space (2001)
• Ted Chamberlin, Gartner analyst (2001)
• David Schaeffer, founder of Cogent Communications (2002)

The Phoenix#

Equinix

• Steve Smith, chief executive (2007–2018), later chief executive of Zayo (2020)
• Lee Theng Kiat, chief executive of ST Telemedia and Equinix's chairman after the rescue (2002)
• Rich Kalbrener, chief executive of Pihana Pacific (2002)

GI Partners and Digital Realty

• Richard Magnuson, founder of GI Partners and executive chairman of Digital Realty (2001–2012)
• Michael F. Foust, chief executive of Digital Realty, later chairman of Vantage (2004–2017)

Switch

• Rob Roy, founder (2002–2022)
• Mike Borden, president, and Rob Roy's brother (2008)

Managed Hosting

• Rob McCormick, chairman and chief executive of Savvis (2004)
• Jim Ousley, chief executive of Savvis under CenturyLink (2011)
• Glen F. Post III, chief executive of CenturyLink (2011)

DuPont Fabros

• Hossein Fateh, chief executive (2008)
• Mark Wetzel, chief financial officer (2008)

The Cloud#

Amazon Web Services

• Jeff Bezos, founder of Amazon (2006)
• Andy Jassy, founding head of Amazon Web Services, who by 2026 was setting Amazon's capital spending (2003–2026)
• Chris MacAskill, of SmugMug, an early AWS customer (2006)

Building Their Own

• Danny Hillis, the computer scientist who called Google's network "the biggest computer in the world" (2006)
• Curt Morris, of the Port of Quincy, and Tim Snead, Quincy's city administrator (2006)
• Mark Zuckerberg, founder of Facebook (2011, 2025)
• Jonathan Heiliger, Facebook's vice president of technical operations (2011)
• Amir Michael, Facebook hardware engineer (2011)

Data Center Alley

• Buddy Rizer, of Loudoun County's economic-development office (2012–2016)
• Ron Meyer, Loudoun County supervisor (2016)

The Wholesale Developers

• Chad Williams, chairman and chief executive of QTS (2013–2021)

Zayo and zColo

• Dan Caruso, founder and chief executive of Zayo Group (2007–2020)
• Greg Friedman, of zColo (2015)
• Kevin Ooley, president of DataBank (2020)
• Raul Martynek, chief executive of DataBank (2020–2021)

The Reconsolidation#

Digital Realty

• Bill Stein, chief executive (2015–2017)
• Christopher Eldredge, chief executive of DuPont Fabros (2017)
• David Ruberg, chief executive of Interxion (2019–2020)

Equinix

• Charles Meyers, chief executive (2018–2019)
• John Hughes, executive chairman of TelecityGroup (2015)

DigitalBridge

• Marc Ganzi, cofounder of Digital Bridge and chief executive of DigitalBridge Group (2013–2022)
• Ben Jenkins, cofounder of Digital Bridge (2013)
• Tom Barrack, founder of Colony Capital (2019)
• Sureel Choksi, chief executive of Vantage (2017)

The Take-Private Wave

• Tyler Henritze and Greg Blank, of Blackstone (2021)
• Dave Ferdman, cofounder and interim chief executive of CyrusOne (2021)
• Bruce Duncan, chief executive of CyrusOne (2020–2021)
• Tom Bartlett, chief executive of American Tower (2021)

The Infrastructure Funds

• Grant van Rooyen, chief executive and chairman of Cologix, and a Level 3 colleague (2017)
• Jan Vesely, of EQT (2019–2020)
• Andrew Schaap, chief executive of Aligned Energy (2018)
• Robin Khuda, founder and chief executive of AirTrunk (2020)
• Frank Kwok, of Macquarie (2020)

The Analyst

• Nick Del Deo, MoffettNathanson analyst who compared dark fiber with towers (2014)

The AI Gold Rush#

Nvidia

• Jensen Huang, cofounder and chief executive (1993–2026)
• Colette Kress, chief financial officer (2025)

OpenAI and Stargate

• Sam Altman, chief executive of OpenAI (2023–2026)
• Masayoshi Son, chairman of SoftBank and of Stargate (2025)
• Larry Ellison, cofounder of Oracle (2025)

CoreWeave

• Michael Intrator, cofounder and chief executive (2017–2026)

xAI

• Elon Musk, founder of xAI and builder of Colossus (2024–2026)

The Cloud Builders

• Amy Hood, chief financial officer of Microsoft (2026)
• Anat Ashkenazi, chief financial officer of Alphabet (2026)
• Rachel Peterson, of Meta (2026)

Power

• Joe Dominguez, chief executive of Constellation Energy (2024)
• Bobby Hollis, of Microsoft (2024)
• Mark Christie, FERC commissioner (2024)
• Scott Strazik, chief executive of GE Vernova (2025)
• Stu Bresler, PJM executive (2025)
• Chris Wright, U.S. Secretary of Energy (2026)

Anthropic

• Dario Amodei, cofounder and chief executive (2021–2026)
• Daniela Amodei, cofounder, president and chair of the board (2021–2026)
• Jaan Tallinn, cofounder of Skype, who led Anthropic's Series A (2021)
• Alfred Lin, of Sequoia Capital (2026)

The Neighbors and the Politicians

• KeShaun Pearson, of Memphis Community Against Pollution (2025)
• Paul Young, mayor of Memphis (2025)
• Abre' Conner, of the NAACP (2026)
• Laura Thoms, of Earthjustice (2026)
• Debra Brown, the federal judge who let the xAI turbine case proceed (2026)
• Greg Abbott, governor of Texas (2025–2026)
• Gina Hinojosa, who defeated Abbott (2026)
• Amy Acton, elected governor of Ohio (2026)
• Greg Casar, U.S. representative (2026)
• Josh Levi and Dan Diorio, of the Data Center Coalition (2024–2026)

The Skeptics and the Analysts

• Jim Covello, head of global equity research at Goldman Sachs (2024)
• David Cahn, of Sequoia Capital (2024)
• Michael Burry, the investor who bet against housing before 2008 (2025)
• Stacy Rasgon, Bernstein analyst (2025)
• Aswath Damodaran, NYU professor who valued Anthropic by its cash flows (2026)

Returning from Bandwidth#

• Warren Buffett, chairman of Berkshire Hathaway, whose Cinderella's Ball frames both books
• Andrew Odlyzko, the AT&T Labs researcher who debunked the Internet traffic myth (1998)
• Jack Grubman, the telecom analyst who amplified it (late 1990s)
• Brad Feld, cofounder of SoftBank's venture fund and author of Bandwidth's foreword (1996–2000)

The sources for each chapter, in the order of the book. Notes in the text explain conflicts and asides; this list is the bibliography.

Chapter I: The Glass House#

• Associated Press. "Company News; I.B.M.-Kodak Pact." New York Times, July 26, 1989.
• Banks, Michael. On the Way to the Web. 2012.
• "Borman Quitting NASA To Join H.R. Perot Firm." Wall Street Journal, January 30, 1970.
• Broy, Anthony. "Computers Aid in Decisions." New York Times, November 14, 1971.
• Business Week. "Special Report." May 24, 1969, 68.
• Carroll, Paul B. "Big Blues: Hurt by a Pricing War, IBM Plans Write-Off And Cut of 10,000 Jobs." Wall Street Journal, December 6, 1989.
• Carroll, Paul B. "IBM Cancels New-Product Announcement --- Embarrassing Move Comes As Rivals Make Inroads In Disk-Drive Business." Wall Street Journal, July 26, 1989.
• Carroll, Paul B., and Hank Gilman. "Big Blues: Mainframe Slowdown And Stiff Competition Put Pressure on IBM." Wall Street Journal, November 23, 1987.
• Carroll, Paul B., and John R. Wilke. "Calculated Move: Computer Firms Find Service Is What Sells, Not Fancier Hardware." Wall Street Journal, August 15, 1989.
• Cole, Robert J. "Arbitrager Had Loss of $100 Million." New York Times, November 3, 1987.
• "Com-Share Reports Financial Condition 'Improved Materially'." Wall Street Journal, July 10, 1970.
• Computerworld, June 21, 1972.
• "Computer Time Sharing Grows Up: Sales Estimated at $140-Million This Year." New York Times, February 18, 1969.
• "Data Transmission Accuses AT&T at FCC." Wall Street Journal, November 13, 1970.
• "Digital Equipment Beats IBM for Kodak Contract." Wall Street Journal, October 26, 1989.
• Electronic Data Systems Corp. Form 10-K for fiscal 2002. Filed March 12, 2003. https://www.sec.gov/Archives/edgar/data/1007456/000093066103001011/d10k.htm.
• "Electronic Data Systems Initial Offer Oversubscribed." Wall Street Journal, September 13, 1968.
• Equinix, Inc. "Form 424(b)(4) Prospectus." Filed August 11, 2000.
• "F.I. duPont Gets Capital Injection Of $15 Million." Wall Street Journal, November 27, 1970.
• "GE Will Offer Time-Sharing System for Its 420 Computer." Wall Street Journal, June 22, 1967.
• "GE's Computer Unit Had $3.1 Million Loss In First Half of '70." Wall Street Journal, August 24, 1970.
• "General Motors Says It Purchased EDS." Wall Street Journal, October 19, 1984.
• "GM Purchase of EDS To Dilute Share Net By 99 Cents for 5 Years." Wall Street Journal, September 24, 1984.
• "Gulf Insurance Bid by University Computing Is Set." Wall Street Journal, August 26, 1968.
• "Heard on the Street." Wall Street Journal, August 7, 1968.
• Holusha, John. "G.M. Entering Computer Field on Major Scale." New York Times, June 29, 1984.
• IBM. "System/360." IBM History, n.d. https://www.ibm.com/history/system-360.
• IBM. "Thomas J. Watson Sr." IBM History, n.d. https://www.ibm.com/history/thomas-watson-sr.
• "IBM Transfers Its Time-Sharing Services To a Subsidiary to Avoid Antitrust." Wall Street Journal, October 23, 1968.
• King, Elliot. Free for All: The Internet's Transformation of Journalism. 2010.
• Koenig, David. "Texas Billionaire and Navy Veteran H. Ross Perot Dies at 89." Associated Press, July 9, 2019. https://www.militarytimes.com/news/pentagon-congress/2019/07/09/texas-billionaire-and-navy-veteran-h-ross-perot-dies-at-89/.
• "Leasco Had Loss Of $30.8 Million In Sept. 30 Year." Wall Street Journal, December 21, 1970.
• Lehner, Urban C., and Laurie P. Cohen. "GM to Acquire EDS in Transaction Valued at as Much as $2.55 Billion." Wall Street Journal, June 29, 1984.
• "LTV, LTV Aerospace, University Computing Sued for $87 Million." Wall Street Journal, February 20, 1970.
• Metz, Robert. "Market Place: Public Relations And Profit Data." New York Times, May 2, 1972.
• Metz, Robert. "Market Place: Western Union And Tender Bid." New York Times, May 9, 1968.
• Miller, Michael W. "Microsoft Public Offering Is Grabbed Up As Company's Stock Promptly Spurts $7." Wall Street Journal, March 14, 1986.
• "New I.B.M. System 360 Can Serve Business, Science and Government." New York Times, April 8, 1964.
• Penn, Stanley. "Computer Competition: Justice Agency Mulls Move to Force IBM Out of the Time-Sharing Service Business." Wall Street Journal, October 1, 1968.
• "Perot Suffers Paper Loss Of $50 Million on Holding." Wall Street Journal, April 21, 1971.
• "Perot's Stock Falls $445-Million in Day." New York Times, April 23, 1970.
• Pollack, Andrew. "Computer Disaster: Business Seeks Antidote." New York Times, August 24, 1983.
• Pollack, Andrew. "I.B.M.'s Aggressive Pricing." New York Times, August 9, 1982.
• Sasser, W. Earl, Christopher W. L. Hart, and James L. Heskett. The Service Management Course: Cases and Readings. 1991.
• Schrage, Michael. "GE Subsidiary to Launch Home Computer Data Network." Washington Post, October 21, 1985.
• Simison, Robert L. "GM Purchase of EDS May Spur Exchange To Alter Its Rules or Delist Auto Maker." Wall Street Journal, July 11, 1984.
• Smith, William D. "Computer Centers Make Gains in Role of Public Utilities." New York Times, April 4, 1965.
• Smith, William D. "Keydata, Time-Sharing's Phoenix." New York Times, August 16, 1970.
• "SMU Mourns Loss of Donor and Leader Charles Wyly." SMU News, August 9, 2011. https://www.smu.edu/news/archives/2011/charles-wyly-dies-09aug2011.
• Standard & Poor's. "CompuServe Inc." Standard OTC Stock Reports, December 14, 1979. Computer History Museum. https://archive.computerhistory.org/resources/access/text/2024/06/102805811-05-01-acc.pdf.
• SunGard Data Systems Inc. Form 8-K, Exhibit 99.1 (slide presentation). Filed October 5, 2004. https://www.sec.gov/Archives/edgar/data/789388/000119312504166645/dex991.htm.
• "Tale of a Tender: How Collins Radio Fended Off Attempt to Take It Over." Wall Street Journal, May 8, 1969.
• "Tymshare Offer Sold Out." Wall Street Journal, September 25, 1970.
• Tuller, David. "Yet Another Way to Profit From the Computer Age." New York Times, August 24, 1986.
• U.S. Department of Justice. "United States' Memorandum in Support of Partial Judgment Termination." United States v. IBM, September 11, 1995. https://www.justice.gov/atr/case-document/united-states-memorandum-support-partial-termination.
• United Press International. "IBM to Design, Manage Data System for Eastman Kodak." July 25, 1989. https://www.upi.com/Archives/1989/07/25/IBM-to-design-manage-data-system-for-Eastman-Kodak/9867617342400/.
• "University Computing Completes $240 Million Gulf Insurance Merger." Wall Street Journal, January 3, 1969.
• "University Computing Division to Be Acquired By Harris-Intertype." Wall Street Journal, August 3, 1972.
• "University Computing Had $38.9 Million Loss In the Third Quarter." Wall Street Journal, November 14, 1972.
• "University Computing Keeps Recent Forecast Of $2.50-a-Share '69 Net." Wall Street Journal, August 29, 1969.
• "University Computing Posts $1.5 Million Net For First Quarter." Wall Street Journal, May 11, 1971.
• Woutat, Donald. "GM Considers Moving 10,000 Employees To EDS Payroll in Proposed Acquisition." Wall Street Journal, September 26, 1984.

Chapter II: Where the Networks Meet#

• Bradner, Scott O. "Re: [ih] Why was MAE East located in Northern Virginia?" Internet History mailing list, August 2021. https://elists.isoc.org/pipermail/internet-history/2021-August/007449.html.
• Buffett, Warren E. "To the Shareholders of Berkshire Hathaway Inc." Berkshire Hathaway 2008 Annual Report, February 27, 2009. https://www.berkshirehathaway.com/letters/2008ltr.pdf.
• Caruso, Dan. Bandwidth: The Untold Story of Ambition, Deception, Innovation That Shaped the Internet Age and the Dot Com Boom. 2024.
• Center for Land Use Interpretation. "8100 Boone Blvd. Data Center, Virginia." CLUI Land Use Database, n.d. https://clui.org/ludb/site/8100-boone-blvd-data-center.
• Corcoran, Elizabeth. "UUNet to Charge For Some Services: Move May Affect Many Internet Service Providers." Washington Post, May 13, 1997.
• CoreSite Realty Corporation. "Form 424(b)(1) Prospectus." Filed September 23, 2010.
• Curran, John. "[ih] Ingrid Burrington on North Virginia." Internet History mailing list, January 11, 2016. https://elists.isoc.org/pipermail/internet-history/2016-January/003697.html.
• Denninger, Karl. "Re: Network Access Points." com-priv mailing list, March 3, 1994. https://diswww.mit.edu/menelaus/com-priv/10658.
• Deutsch, Claudia H. "On Electronic Highway, Manhattan Is a Destination." New York Times, July 23, 1995.
• "Developers Gear Up for New Telecom Hotels — Buildings Become Centers for Switches and Servers, But Some See a Glut." Wall Street Journal, March 8, 2000.
• Dunlap, David W. "Commercial Real Estate; Once and Future Telecommunications Crossroads." New York Times, January 19, 2000.
• Fry, Jason. "Behind the Scenes — The Answer Man." Wall Street Journal, December 8, 1997.
• Haggerty, Maryann. "AboveNet Picks Reston as Site of 'One-Hop Network.'" Washington Post, July 5, 1999.
• Holusha, John. "Commercial Property/111 Eighth Avenue; From Back Office to Front Office and High Tech." New York Times, January 25, 1998.
• "Is the Internet Outgrowing Its Volunteer Traffic Cops? — 'Root Servers' Need a Paid Staff, Some Critics Say." Wall Street Journal, September 12, 1997.
• "L.A. Looks to Limit Telco Hotels." Wall Street Journal, August 2, 2000.
• Lewis, Peter H. "U.S. Begins Privatizing Internet's Operations." New York Times, October 24, 1994.
• Loudoun County Economic Development. "MAE-East Internet Exchange Point Moves to Loudoun." Timeline entry, n.d. https://biz.loudoun.gov/timeline/mae-east-internet-exchange-point-moves-to-loudoun.
• Markoff, John. "Old Man Bandwidth; Will Commerce Flourish Where Rivers of Wire Converge?" New York Times, December 8, 1997.
• Meeker, Mary, and Chris DePuy. "The Internet Report." Morgan Stanley, February 1996.
• "MFS, Buying UUNet, Intends to Double Its Capital Spending." Wall Street Journal, May 8, 1996.
• National Science Foundation. Award CNS-9321077, "Network Access Point Manager for NSFNET and the NREN." Principal investigator Steve Feldman, MFS Datanet Inc., 1994. https://www.nsf.gov/awardsearch/show-award/?AWD_ID=9321077.
• National Science Foundation. "NSF 93-52: Network Access Point Manager, Routing Arbiter, Regional Network Providers, and Very High Speed Backbone Network Services Provider for NSFNET and the NREN(SM) Program." Program solicitation, May 6, 1993. https://nsarchive.gwu.edu/media/19098/ocr.
• "On-Line Users Attracted by Low Rates Strain System, Causing Frustration." Wall Street Journal, December 18, 1996.
• Pae, Peter, and Rajiv Chandrasekaran. "Major Firm Plans Move to Loudoun: Internet Company to Employ 3,000." Washington Post, December 24, 1997.
• "Space for Rent: Fully Wired, Windows Optional — Developers Seek New Life for 'Telecom Hotels' As Buildings Sit Empty." Wall Street Journal, June 20, 2001.
• Stamler, Bernard. "Neighborhood Report: New York Noise; Instead of the Singing Telegram, Generators and Cooling Systems." New York Times, October 24, 1999.
• Sugawara, Sandra. "Washington Business: On-Line and on the Move: Vienna's Computer Service Attracts the Attention of Consumers, Competitors." Washington Post, August 8, 1994.
• Swisher, Kara. "Anticipating the Internet: Good Timing, Good Deal-Making and Good Luck Turned Rick Adams's UUNet Into a Star." Washington Post, May 6, 1996.
• "Web Crunch: Slow Crawl on the Internet: Massive Amounts of Data Clog Systems Studded With Roadblocks." Wall Street Journal, August 23, 1996.
• Wolff, Stephen. "Network Access Points." com-priv mailing list, March 2, 1994. https://diswww.mit.edu/menelaus/com-priv/10630.
• Woodcock, Bill. "[ih] Why was MAE East located in Northern Virginia?" Internet History mailing list, August 31, 2021. https://elists.isoc.org/pipermail/internet-history/2021-August/007450.html.
• "WorldCom to Update UUNet's Network, Investing $300 Million." Wall Street Journal, February 20, 1997.

Chapter III: The Hosting Gold Rush#

• "Analyst Helps Spark Internet Rally." InternetNews.com, March 26, 1999. https://www.internetnews.com/it-management/analyst-helps-spark-internet-rally.
• Associated Press. "Akamai Technologies Stock Surges on First Trading Day." New York Times, October 30, 1999.
• Auerbach, Jon G. "The Providers — Playing the New Order." Wall Street Journal, November 15, 1999.
• Blair, Jayson. "There's Money in Housing All Those Internet Servers, but There's a Power Drain, Too." New York Times, April 8, 2001.
• Bloomberg News. "Settlement in WorldCom Deal Is Approved." New York Times, April 7, 2001.
• Bloomberg News. "WorldCom Said to Agree on Intermedia Purchase for $39 a Share." New York Times, September 5, 2000.
• Blumenstein, Rebecca, and Jon G. Auerbach. "Qwest, IBM Plan to Build Internet Data Centers." Wall Street Journal, March 27, 2000.
• "Business Brief — CMGI: Dell, Microsoft Purchase Minority Stakes in NaviSite." Wall Street Journal, June 8, 1999.
• "Business Brief — Data Return Corp.: IPO Is Scheduled Today; Shares Are Priced at $13." Wall Street Journal, October 28, 1999.
• "Business Brief — Equinix Inc.: Plan Is Set for Exchanges To Manage Internet Traffic." Wall Street Journal, May 26, 1999.
• "Business Brief — Microsoft Corp.: Investment of $5 Million Is Made in Data Return." Wall Street Journal, November 3, 1999.
• "Business Brief — WorldCom Inc.: Court Lets Intermedia Deal Proceed, Warns of Damages." Wall Street Journal, December 14, 2000.
• Carlton, Jim. "Exodus Finds 'Web Hosting' May Be Profitable Niche." Wall Street Journal, April 23, 1998.
• Carlton, Jim. "Venerable Builder Bechtel Gets a New Lift From the Dot-Com Boom." Wall Street Journal, March 1, 2000.
• Caruso, Dan. Bandwidth: The Untold Story of Ambition, Deception, Innovation That Shaped the Internet Age and the Dot Com Boom. 2024.
• Cauley, Leslie. "Former Chief of Cable TV at AT&T To Be CEO of Global Crossing Unit." Wall Street Journal, December 7, 1999.
• Clark, Don. "E-Business: Starting Gate." Wall Street Journal, February 7, 2000.
• Clark, Don. "South Asian 'Angels' Reap Riches, Spread Wealth in Silicon Valley." Wall Street Journal, May 2, 2000.
• "Dell Forms Unit to Market to Web Hosts, Services." Wall Street Journal, February 7, 2000.
• "Deals & Deal Makers: IPO Scorecard." Wall Street Journal, August 17, 2000.
• Deogun, Nikhil, Anita Raghavan, and Deborah Solomon. "Global Crossing Weighs Sale of 2 Units." Wall Street Journal, July 12, 2000.
• Deogun, Nikhil, and Deborah Solomon. "Exodus Restarts Talks for Global Crossing Unit." Wall Street Journal, September 28, 2000.
• Deogun, Nikhil, and Deborah Solomon. "Global Crossing's Talks on Sale Of Unit to Exodus Stall on Price." Wall Street Journal, July 17, 2000.
• "Digex Shareholder Gets Judge to Expedite Suit To Stop Intermedia Sale." Wall Street Journal, October 5, 2000.
• "Equinix Inc. to Unveil Large Finance Deal For Internet Centers." Wall Street Journal, December 2, 1999.
• Equinix, Inc. Prospectus, Rule 424(b)(4). Filed August 11, 2000.
• "Equinix to Construct 30 Exchange Centers For Internet Traffic." Wall Street Journal, November 4, 1999.
• "Exodus and At Home Will Jointly Market Web Hosting Services." Wall Street Journal, March 27, 1998.
• Exodus Communications, Inc. "Form 10-K405 Annual Report for the Fiscal Year Ended December 31, 1998." Filed February 22, 1999.
• Exodus Communications, Inc. "Form 10-K Annual Report for the Fiscal Year Ended December 31, 1999." Filed March 30, 2000.
• Exodus Communications, Inc. "Form DEFM14A Definitive Proxy Statement (Merger with GlobalCenter Holding Co.)." Filed December 7, 2000.
• Exodus Communications. Form 10-K for fiscal 2000. Filed March 30, 2001.
• Exodus Communications. Prospectus, Rule 424(b)(4). Filed March 19, 1998.
• Flaherty, Julie. "Private Sector; When a 'Hit' Is a Grand Slam." New York Times, December 5, 1999.
• Globix Corp. Form 10-K for fiscal 2001. Filed December 31, 2001. https://www.sec.gov/Archives/edgar/data/1003111/000095013001506317/d10k.txt.
• Goodman, Peter S., and Renae Merle. "End of Its Merger Run Led to WorldCom's Fall." Washington Post, June 30, 2002.
• Grimes, Ann. "Exodus to Buy Unit Of Global Crossing In Deal for Stock." Wall Street Journal, September 29, 2000.
• Harris, Nicole. "Sprint's Program for Web Hosting Could Possibly Include a Purchase." Wall Street Journal, December 12, 2000.
• Hennessey, Raymond. "Deals & Deal Makers: Genuity Falls 14% In Market Debut, Weighted by Losses." Wall Street Journal, June 29, 2000.
• Hennessey, Raymond. "Deals & Deal Makers — IPO Outlook: Heavy IPO Calendar Affecting Offerings." Wall Street Journal, August 14, 2000.
• Holusha, John. "Commercial Property/High Technology; A 500,000-Sq.-Ft. Lease Brings Qwest to Yonkers." New York Times, November 5, 2000.
• Holusha, John. "Commercial Property; Web Gives White Elephants a New Life." New York Times, July 9, 2000.
• Holusha, John. "Commercial Real Estate; Plans for a 'Server Farm' Take Developer Upstate." New York Times, September 6, 2000.
• Huber, Peter, and Mark P. Mills. "Dig More Coal—the PCs Are Coming." Forbes, May 31, 1999.
• Hymowitz, Carol. "In the Lead: Fired From Two Jobs, A High-Tech Manager Gets the Last Laugh." Wall Street Journal, October 5, 1999.
• In re Digex Inc. Shareholders Litigation, 789 A.2d 1176 (Del. Ch. 2000). Decided December 13, 2000. Exhibit 99.1 to Intermedia Communications Form 8-K. https://www.sec.gov/Archives/edgar/data/885067/000095012300011575/0000950123-00-011575.txt.
• "Intel to Shutter Its Operation For Web Hosting." Wall Street Journal, June 19, 2002.
• Lawrence Berkeley National Laboratory. "Data Center Energy Use: Truth versus Myth." Science Beat, July 16, 2002. https://www2.lbl.gov/Science-Articles/Archive/data-center-energy-myth.html.
• Lohr, Steve, with John Markoff. "Computing Centers Become the Keeper of Web's Future." New York Times, May 19, 1999.
• McWilliams, Gary. "Dell to Sell Assets Of Web-Host Unit To Sprint Corp." Wall Street Journal, August 31, 2001.
• Mehta, Stephanie N. "Metromedia Fiber Network Plans to Buy AboveNet for $1.76 Billion in Stock." Wall Street Journal, June 24, 1999.
• Metromedia Fiber Network, Inc. Form 10-Q/A. Filed January 26, 2001. https://www.sec.gov/Archives/edgar/data/1043533/000091205701003042/a2036114z10-qa.txt.
• "Navisite to Reduce Its Staff." Wall Street Journal, February 6, 2001.
• NaviSite, Inc. Form 10-K for fiscal 2001. Filed October 30, 2001. https://www.sec.gov/Archives/edgar/data/1084750/000092701601503388/d10k.txt.
• Peck, Georgia. "Louisiana Public Service Commission Green Lights Entergy Power Plant Construction for Meta Data Center." WAFB, August 20, 2025. https://www.wafb.com/2025/08/20/louisiana-public-service-commission-green-lights-entergy-power-plant-construction-meta-data-center/.
• Perez, Evan. "Internet Hub May Revive Miami Area." Wall Street Journal, June 28, 2000.
• Plotkin, Hal. "Exodus's Stock Seen Going Higher." CNBC.com, January 2000. https://plotkin.com/?p=767.
• Pollack, Andrew. "Web-Page Distribution System Could Unclog Internet Traffic Jams." New York Times, April 26, 1999.
• "Professor Stands to Reap Windfall in Akamai IPO." Wall Street Journal, October 26, 1999.
• Reuters. "Company News; Compaq to Sell an Internet Services Unit to AboveNet." New York Times, May 25, 1999.
• Reuters. "Company News; Metromedia to Buy AboveNet for $1.55 Billion." New York Times, June 24, 1999.
• Reuters. "Exodus Is Buying Cohesive Technology for $100 Million." New York Times, April 23, 1999.
• Rich, Motoko. "Developers Gear Up for New Telecom Hotels — Buildings Become Centers For Switches and Servers, But Some See a Glut." Wall Street Journal, March 8, 2000.
• Schiesel, Seth. "Daniel Lewin, Technology Executive, 31." New York Times, September 13, 2001.
• Schiesel, Seth. "Technology; Genuity to Introduce First Major Internet Service Offerings." New York Times, September 19, 2000.
• Solomon, Deborah. "Digex Inc. Shares Decline Almost 20% As WorldCom Confirms Intermedia Bid." Wall Street Journal, September 6, 2000.
• Solomon, Deborah. "Genuity Unit Of GTE Plans Stock Offering." Wall Street Journal, May 25, 2000.
• Solomon, Deborah. "WorldCom, Intermedia Agree to Revise Terms of Deal in Bid to Settle Digex Suit." Wall Street Journal, February 16, 2001.
• Solomon, Deborah, and Nikhil Deogun. "Deals & Deal Makers: Digex Sale Reflects Minority Holders' Secondary Status." Wall Street Journal, September 8, 2000.
• Solomon, Deborah, and Nikhil Deogun. "WorldCom to Acquire Intermedia for Stock, Debt." Wall Street Journal, September 5, 2000.
• Templin, Neal. "Power-Hungry Web 'Server Farms' Find Cooler Reception in California." Wall Street Journal, February 28, 2001.
• "Technology Brief — Exodus Communications Inc.: Purchase of Service Metrics In Stock Deal Is Planned." Wall Street Journal, October 28, 1999.
• Terremark Worldwide, Inc. Form 10-K for fiscal year ended March 31, 2001. Filed July 16, 2001. https://www.sec.gov/Archives/edgar/data/912890/000095014401504501/g70341e10-k.txt.
• "The Markets; Deal of the Day." New York Times, March 20, 1998.
• U.S. House of Representatives, Committee on Government Reform, Subcommittee on National Economic Growth, Natural Resources, and Regulatory Affairs. Kyoto and the Internet: The Energy Implications of the Digital Economy. Hearing, 106th Cong., 2nd sess., February 2, 2000. Serial No. 106-125.
• U.S. House of Representatives, Committee on Oversight and Government Reform, Subcommittee on Economic Growth, Energy Policy, and Regulatory Affairs. America's AI Moonshot: The Economics of AI, Data Centers, and Power Consumption. Hearing, 119th Cong., 1st sess., April 1, 2025. Serial No. 119-15.
• WorldCom, Inc. "Form S-4 Registration Statement." Filed October 16, 2000.
• WorldCom, Inc. "Form S-4/A Amendment to Registration Statement." Filed April 26, 2001.
• Young, Shawn. "WorldCom's Intermedia Deal Hits Bump." Wall Street Journal, December 15, 2000.
• Zuckerman, Gregory. "A Year After the Peak: How Nasdaq's Mighty Have Fallen." Wall Street Journal, March 5, 2001.

Chapter IV: Empty Floors#

• Associated Press. "Exodus Communications Files for Chapter 11." Wall Street Journal, September 27, 2001.
• Associated Press. "Technology; Exodus Chief Unexpectedly Resigns Post." New York Times, September 5, 2001.
• Berman, Dennis K., Phillip Day, and Henry Sender. "Global Crossing Files for Chapter 11, Plans to Reorganize with Asia Firms." Wall Street Journal, January 29, 2002.
• Bloomberg News. "Settlement in WorldCom Deal Is Approved." New York Times, April 7, 2001.
• Bloomberg News. "Technology Briefing | Internet: Globix Plans to Seek Protection." New York Times, January 15, 2002.
• "Business Brief — Divine Inc.: Data Return Corp. Purchase Is Planned in $31 Million Deal." Wall Street Journal, November 5, 2001.
• Caruso, Dan. Bandwidth: The Untold Story of Ambition, Deception, Innovation That Shaped the Internet Age and the Dot Com Boom. 2024.
• Coffman, K. G., and A. M. Odlyzko. "Growth of the Internet." AT&T Labs–Research, July 6, 2001. https://www-users.cse.umn.edu/~odlyzko/doc/oft.internet.growth.pdf.
• Cogent Communications Group, Inc. "Form 8-K/A." Filed June 17, 2002.
• "Digex Inc. Cuts 7% of Work Force." Wall Street Journal, June 28, 2002.
• Digital Realty Trust, Inc. "Form 424(b)(4) Prospectus." Filed November 1, 2004.
• DuPont Fabros Technology, Inc. "Form 424(b)(4) Prospectus." Filed October 19, 2007.
• eWEEK Editors. "Analyst: Exodus Bankruptcy Imminent." eWeek, September 25, 2001. https://www.eweek.com/news/analyst-exodus-bankruptcy-imminent/.
• EXDS, Inc. "Second Amended Disclosure Statement with Respect to Second Amended Joint Plan of Reorganization of EXDS, Inc. and Its Debtor Affiliates." Case No. 01-10539 (SLR), U.S. District Court for the District of Delaware. Exhibit to Form 8-K. Filed May 3, 2002.
• Exodus Communications, Inc. "Form 8-K." Filed September 27, 2001.
• Garbarine, Rachelle. "In the Region/New Jersey; A Few Telco Hotels Are Still in the Works." New York Times, May 6, 2001.
• Genuity Inc. "Genuity Announces Tender Offer to Acquire Integra." Form 8-K, Exhibit 99.1. Filed June 1, 2001. https://www.sec.gov/Archives/edgar/data/1110794/000092701601501262/dex991.txt.
• "Genuity Is Left on the Launching Pad." Forbes, July 26, 2002. https://www.forbes.com/2002/07/26/0726genuity.html.
• Globix Corporation. "Form 8-K (Event Date March 1, 2002)."
• Globix Corporation. "Form 8-K (Event Date April 8, 2002)."
• Globix Corporation. "Form 10-Q." Filed May 15, 2002.
• Goodman, Peter S., and Renae Merle. "End of Its Merger Run Led to WorldCom's Fall." Washington Post, June 30, 2002.
• Hunt, Ben. "Cable and Wireless in Deal to Acquire Exodus for $850 Million." Financial Times, via New York Times, November 30, 2001.
• Mangalindan, Mylene. "E-Business: Exodus Suits Up to Stay in Game." Wall Street Journal, August 27, 2001.
• Mangalindan, Mylene. "E-Business: Overcapacity Jitters Hurt Web-Hosting Stocks." Wall Street Journal, June 18, 2001.
• Mangalindan, Mylene. "Exodus Communications Readies A Filing Under Bankruptcy Code." Wall Street Journal, September 25, 2001.
• Mangalindan, Mylene. "Exodus's Chairman Hancock Resigns; Board Names Krause as Her Successor." Wall Street Journal, September 5, 2001.
• Manuel, Gren. "Cable & Wireless Establishes Foothold, Faces Obstacles in Global Web-Hosting War." Wall Street Journal, December 3, 2001.
• McWilliams, Gary. "Dell to Sell Assets Of Web-Host Unit To Sprint Corp." Wall Street Journal, August 31, 2001.
• Miller, Rich. "Digital Realty Acquires AboveNet Centers." DataCenter Knowledge, October 2, 2006.
• "Net-Services Firm Genuity to Buy France's Integra for $113.7 Million." Wall Street Journal, June 1, 2001.
• Noguchi, Yuki. "Cogent to Buy U.S. Division of PSINet." Washington Post, February 27, 2002.
• Nuzum, Christine. "Metromedia Fiber Files for Bankruptcy Protection." Wall Street Journal, May 20, 2002.
• Pacelle, Mitchell, and Mylene Mangalindan. "Cable & Wireless to Buy Most of Exodus For $570 Million, Plus Debt Assumption." Wall Street Journal, November 30, 2001.
• "PSINet Reduces Founder's Role." InternetNews, April 30, 2001. https://internetnews.com/archive/psinet-reduces-founders-role.
• PSINet Inc. Form 8-K. Filed June 1, 2001. https://www.sec.gov/Archives/edgar/data/940716/000090130901500044/form8k_june12001.txt.
• Schiesel, Seth. "Daniel Lewin, Technology Executive, 31." New York Times, September 13, 2001.
• Scinta, Christopher. "Metromedia Fiber Sees Enterprise Value of $230M–$330M." Wall Street Journal, May 22, 2003.
• Sender, Henny. "Telecom Special: In Bankruptcy, PSINet Gets '90% Off' Tag." Wall Street Journal, November 27, 2001.
• Simonson, Sharon. "Equinix Buys AboveNet's Data Center in San Jose." Silicon Valley Business Journal, December 12, 2004.
• Smith, Ray A. "Space for Rent: Fully Wired, Windows Optional — Developers Seek New Life For 'Telecom Hotels' As Buildings Sit Empty." Wall Street Journal, June 20, 2001.
• Smith, Ray A. "'Telecom Hotels' Find Interest Once More." Wall Street Journal, October 10, 2001.
• Templin, Neal. "Power-Hungry Web 'Server Farms' Find Cooler Reception in California." Wall Street Journal, February 28, 2001.
• Tier 1 Research. "Infrastructure Bible v2.0C: Where Has All the Space Gone?" Tier 1 Research, April 2002. Press release, April 4, 2002. https://www.whtop.com/news/13496-tier-1-releases-hosting-directory-data-center-report.
• "Verizon Global, Genuity Solutions Cancel Metromedia Fiber Pacts." Wall Street Journal, April 7, 2002.
• Zuckerman, Gregory. "A Year After the Peak: How Nasdaq's Mighty Have Fallen." Wall Street Journal, March 5, 2001.

Chapter V: The Phoenix#

• Amazon Web Services. "Announcing Amazon Elastic Compute Cloud (Amazon EC2) - beta." AWS What's New, August 24, 2006.
• Associated Press. "Verizon to Buy Terremark for $1.4 Billion." Phys.org, January 28, 2011.
• Caruso, Dan. Bandwidth: The Untold Story of Ambition, Deception, Innovation That Shaped the Internet Age and the Dot Com Boom. 2024.
• CenturyLink and Savvis. "CenturyLink to Acquire Savvis for $40 Per Share in Cash and Stock Transaction." Press release, Savvis Form 8-K, Exhibit 99.2. April 27, 2011.
• CenturyLink. "CenturyLink and Savvis Complete Merger." Press release, Form 8-K, Exhibit 99.1. July 15, 2011.
• Digital Realty Trust. Form 10-K for fiscal 2004. Filed 2005 (accession 0001193125-05-067198).
• Digital Realty Trust. Prospectus, Rule 424(b)(4). October 28, 2004.
• Equinix and Switch and Data. "Equinix to Acquire Switch and Data in $689 Million Deal That Expands Equinix's Presence Into 16 New Markets." Form 425. October 21, 2009.
• "Equinix Completes Switch & Data Merger." Data Center Knowledge, May 3, 2010.
• Equinix, Inc. "Amendment No. 1 to Form S-4 Registration Statement" (proxy statement/prospectus for the Switch & Data Facilities Company special meeting). Filed December 21, 2009.
• Equinix. "Equinix Completes $683.4 Million Acquisition of Switch and Data, Strengthening Its Leadership in Data Center Services." Press release, Form 8-K, Exhibit 99.1. May 3, 2010.
• Equinix. "Equinix Completes Merger and Financing Transactions." Press release, Form 8-K, Exhibit 99.1. December 31, 2002.
• Equinix. "Equinix Names Former HP Executive, Stephen M. Smith, CEO." Press release. March 19, 2007.
• Equinix. "Equinix Reports Fourth Quarter and Year-End 2008 Results." Press release. February 11, 2009.
• Equinix. "Equinix to Open Third Data Center in the Silicon Valley; Company Updates 2005 Guidance." Press release, Form 8-K, Exhibit 99.1. December 8, 2004.
• Equinix. Form 10-K for fiscal 2003. Filed 2004.
• Equinix. Form 10-K for fiscal 2004. Filed 2005.
• Equinix. Form 8-K. Date of report April 30, 2010 (accession 0001193125-10-104230).
• Equinix. Form 8-K. Filed December 2004 (accession 0001193125-04-211084).
• Equinix. Form 8-K. Filed October 9, 2002.
• EXDS, Inc. (f/k/a Exodus Communications, Inc.). "Second Amended Disclosure Statement with Respect to Second Amended Joint Plan of Reorganization of EXDS, Inc. and Its Debtor Affiliates." Case No. 01-10539 (SLR), U.S. District Court for the District of Delaware, April 2002. Exhibit to EXDS Form 8-K, filed May 3, 2002.
• GI Partners. "GI Partners to Manage CalPERS' $3.4 Billion CalEast Industrial and Logistics Real Estate Portfolio." Press release. December 1, 2010.
• "Jeff Bezos' Risky Bet." BusinessWeek, November 13, 2006.
• Kerner, Sean Michael. "Court Approves Savvis Bid for Bankrupt C&W." InternetNews, January 23, 2004.
• Latamore, Bert. "Switch and the Super|Nap: The Story of the World's Highest Density Data Center." SiliconANGLE, September 21, 2010.
• "Legislators told data center tax help will benefit Nevada big." Las Vegas Review-Journal, 2015.
• Miller, Rich. "2009: Stellar Year for Data Center Stocks." Data Center Knowledge, January 4, 2010.
• Miller, Rich. "Analysts: Demand Still Outpacing Supply." Data Center Knowledge, June 28, 2010.
• Miller, Rich. "Digital Realty in 'A Different World'." Data Center Knowledge, March 2, 2009.
• Miller, Rich. "DuPont Fabros Halts Data Center Projects." Data Center Knowledge, November 17, 2008.
• Miller, Rich. "DuPont Fabros Seeks Funding, Shares Plunge." Data Center Knowledge, November 6, 2008.
• Miller, Rich. "Equinix Acquires Fourth Silicon Valley Data Center." Data Center Knowledge, June 27, 2005.
• Miller, Rich. "Equinix Buys IXEurope for $482M." Data Center Knowledge, June 28, 2007.
• Miller, Rich. "GI Partners Sells Last Data Center Holding." Data Center Knowledge, September 14, 2009.
• Miller, Rich. "Magnuson to Step Down as Digital Realty Chairman." Data Center Knowledge, March 9, 2012.
• Miller, Rich. "Rackspace IPO Priced at $12.50." Data Center Knowledge, August 8, 2008.
• Miller, Rich. "Terremark: Credit Crunch Scuttled Potential Sale." Data Center Knowledge, November 5, 2008.
• Miller, Rich. "The Vegas SuperNAP: A Data Center Revolution?" Data Center Knowledge, May 27, 2008.
• Miller, Rich. "Tier 1: Oversupply Fears Are Unfounded." Data Center Knowledge, September 27, 2011.
• Plaskon, Ky. "Profiting from Enron Bankruptcy." KNPR (Nevada Public Radio), November 29, 2004.
• Rackspace Hosting. Prospectus, Rule 424(b)(1). August 7, 2008.
• "Savvis Acquires C&W US Assets." Light Reading, January 23, 2004.
• SAVVIS, Inc. Definitive Proxy Statement (Schedule 14A), Merger with CenturyLink. Filed June 10, 2011.
• Savvis. "SAVVIS Closes Purchase of Cable & Wireless America Assets." Press release, Form 8-K, Exhibit 99.1. March 8, 2004.
• Switch and Data. "Switch and Data Reports Third Quarter 2009 Financial and Operating Results." Press release, Form 8-K, Exhibit 99.1. October 27, 2009.
• Switch, Inc. Prospectus, Rule 424(b)(4). Filed October 10, 2017.
• "Telecomeback." Forbes, January 21, 2002.
• Terremark Worldwide, Inc. Solicitation/Recommendation Statement on Schedule 14D-9. Filed February 11, 2011.
• Terremark Worldwide. Form 10-K for fiscal year ended March 31, 2001. Filed 2001.
• Vance, Ashlee. "Welcome to Las Vegas – Home of the Technology Superpower You've Never Heard Of." The Register, May 24, 2008.
• Vaughn, Alan. "Pihana Pacific Enters Multinational Merger." Honolulu Star-Bulletin, October 3, 2002.
• Verizon. "Verizon Completes Terremark Acquisition." Press release, Form 8-K, Exhibit 99.1. April 11, 2011.
• Verizon. "Verizon to Acquire Terremark, Boosting Cloud Strategy through Powerful Unified Enterprise IT Delivery Platform." Press release, Terremark Form 8-K, Exhibit 99.2. January 27, 2011.
• "Witness: Enron Broadband Software 'Pixie Dust'." NBC News, April 21, 2005.

Chapter VI: The Cloud Eats the Building#

• "After dramatic growth, Ashburn expects even more data centers." Washington Post, Capital Business, August 2011.
• Amazon Web Services. "Announcing Amazon Elastic Compute Cloud (Amazon EC2) - beta." AWS What's New, August 24, 2006.
• Amazon.com. "Amazon Web Services Launches." Press release, Business Wire, March 14, 2006.
• Associated Press. "High-tech firms move into potato country." NBC News, March 28, 2006.
• Bell, Eric. "DataBank acquires zColo sites from Zayo for $1.4 Billion." Baxtel, September 30, 2020.
• Caruso, Dan. Bandwidth: The Untold Story of Ambition, Deception, Innovation That Shaped the Internet Age and the Dot Com Boom. 2024.
• Cincinnati Bell Inc. Form 8-K. Filed May 13, 2010.
• Cincinnati Bell. "CyrusOne Inc. Prices Initial Public Offering of Common Stock." Exhibit 99.1 to Form 8-K, January 17, 2013.
• Colony Capital, Inc. Form 10-K for fiscal 2020. Filed March 1, 2021.
• CoreSite Realty Corporation. Prospectus, Rule 424(b)(1). Filed September 23, 2010.
• CyrusOne Inc. Prospectus, Rule 424(b)(4). Filed January 22, 2013.
• DataBank. "DataBank Completes zColo Transaction with Closing Acquisition of Five Data Centers in France." PR Newswire, February 3, 2021.
• Del Deo, Nick, and Sydney Marks. "Zayo: Plan A, Version 2.0… A Closer Look at zColo." MoffettNathanson, March 4, 2019.
• Del Deo, Nick. "Data Centers: Implications of the Hyperscale Leasing Slowdown and a Close Look at Valuation." MoffettNathanson, July 22, 2019.
• Del Deo, Nick. "U.S. Communications Infrastructure: Meeting with Dan Caruso, Founder and Former Chairman & CEO of Zayo." MoffettNathanson, February 20, 2025.
• Del Deo, Nick. "Zayo: Q2 F2020 Earnings: Good Night, and Good Luck." MoffettNathanson, February 4, 2020.
• DuPont Fabros Technology, Inc. Prospectus, Rule 424(b)(4). Filed October 19, 2007.
• Equinix. "Equinix Board of Directors Approves REIT Conversion." Press release, December 23, 2014.
• Facebook. "Facebook Launches Open Compute Project." Press release, April 7, 2011.
• Freed, Benjamin. "70 Percent of the World's Web Traffic Flows Through Loudoun County." Washingtonian, September 14, 2016.
• Glanz, James. "Data Barns in a Farm Town, Gobbling Power and Flexing Muscle." New York Times, September 24, 2012.
• "Google buying 30 acres on Columbia River." GlobeSt, February 18, 2005.
• "Jeff Bezos' Risky Bet." BusinessWeek, November 13, 2006.
• Markoff, John, and Saul Hansell. "Hiding in Plain Sight, Google Seeks More Power." New York Times, June 14, 2006. Reprinted in Infowarrior mailing list, June 2006. https://attrition.org/pipermail/infowarrior/2006-June/000318.html.
• McWilliams, John, and Matthew Nevinger. "Data Center Spotlight: Power and Lease Pricing Outlook." Cushman & Wakefield, July 26, 2023.
• Miller, Rich. "2009: Stellar Year for Data Center Stocks." Data Center Knowledge, January 4, 2010.
• Miller, Rich. "CoreSite Goes Public With IPO on NYSE." Data Center Knowledge, September 23, 2010.
• Miller, Rich. "Data Center Leasing in Northern Virginia Hits Record 115 MW in 2017." Data Center Frontier, February 2, 2018.
• Miller, Rich. "DataBank Acquires zColo Data Centers in Major Expansion." Data Center Frontier, September 29, 2020.
• Miller, Rich. "DuPont Fabros Halts Data Center Projects." Data Center Knowledge, November 17, 2008.
• Miller, Rich. "DuPont Fabros Raises $641M in IPO." Data Center Knowledge, October 19, 2007.
• Miller, Rich. "DuPont Fabros Seeks Funding, Shares Plunge." Data Center Knowledge, November 6, 2008.
• Miller, Rich. "Facebook Has Spent $210 Million on Oregon Data Center." Data Center Knowledge, January 30, 2012.
• Miller, Rich. "Facebook Opens its Server, Data Center Designs." Data Center Knowledge, April 7, 2011.
• Miller, Rich. "Loudoun: 5 Million Square Feet of Data Centers, More to Come." Data Center Knowledge, September 11, 2012.
• Miller, Rich. "WashPost Profiles Ashburn 'Data Center Alley.'" Data Center Knowledge, August 28, 2011.
• Miller, Rich. "WSJ: Google is Buying 111 8th Avenue." Data Center Knowledge, December 3, 2010.
• Miller, Ron. "How AWS came to be." TechCrunch, July 2, 2016.
• Peterson, Tom. "First Google Data Center hits tax rolls; Bill just went up $5 Million." Columbia Community Connection, October 2022.
• QTS Realty Trust, Inc. Prospectus, Rule 424(b)(4). Filed October 10, 2013.
• QTS Realty Trust. "QTS Realty Trust, Inc. Prices Initial Public Offering." Press release, PR Newswire, October 9, 2013.
• RebornData. "Amazon Betas 'Elastic' Grid Computing Service." Slashdot, August 24, 2006.
• Renaissance Capital. "QTS Realty Trust prices IPO at $21, below the range." October 9, 2013.
• Rogoway, Mike. "Google buys more land in The Dalles." The Oregonian, reprinted in Spokesman-Review, March 29, 2016.
• Shehabi, Arman, Sarah Smith, Dale Sartor, Richard Brown, Magnus Herrlin, Jonathan Koomey, Eric Masanet, Nathaniel Horner, Inês Azevedo, and William Lintner. United States Data Center Energy Usage Report. LBNL-1005775. Berkeley, CA: Lawrence Berkeley National Laboratory, June 2016.
• Silver Lake. "Vantage Data Centers Announces Launch and Funding from Silver Lake." Press release, June 17, 2010.
• Sverdlik, Yevgeniy. "Its zColo Acquisition Will Make DataBank a US Data Center Heavyweight." Data Center Knowledge, September 30, 2020.
• Sverdlik, Yevgeniy. "Why Zayo is Paying $675M for Latisys." Data Center Knowledge, January 15, 2015.
• U.S. Congress. Public Law 109-431 (H.R. 5646). December 20, 2006.
• U.S. Environmental Protection Agency, ENERGY STAR Program. Report to Congress on Server and Data Center Energy Efficiency, Public Law 109-431. August 2, 2007.
• Uptime Institute. 2018 Global Data Center Survey. Keynote Report 13, June 2018.
• Zayo Group Holdings, Inc. Definitive Proxy Statement (Schedule 14A), Merger with Front Range TopCo. Filed June 26, 2019.
• Zayo Group Holdings, Inc. Form 10-K for fiscal year ended June 30, 2015. Filed September 18, 2015.
• Zayo Group Holdings, Inc. Form 10-K for fiscal year ended June 30, 2019. Filed September 4, 2019.
• Zayo Group, LLC. Form 10-K for fiscal year ended June 30, 2011. Filed September 9, 2011.
• Zayo Group, LLC. Form 10-K for fiscal year ended June 30, 2013. Filed September 23, 2013.
• Zayo Group. "Zayo Completes Sale of zColo's US and UK Data Center Assets to DataBank." Business Wire, December 15, 2020.

Chapter VII: The Reconsolidation#

• American Tower and CoreSite. "American Tower to Acquire CoreSite." Business Wire release via TelecomTV, November 15, 2021.
• American Tower Corporation. Form 10-Q for the quarter ended June 30, 2022. Filed with the SEC, 2022.
• American Tower. "American Tower Completes Acquisition of CoreSite Realty Corporation." Press release, Exhibit 99.2 to Form 8-K, December 28, 2021.
• "AT&T sells 31 data centers to Brookfield Infrastructure for $1.1 billion." Converge Digest, January 2, 2019.
• AT&T. "AT&T Closes Sale of Data Center Colocation Operations and Assets to Brookfield Infrastructure." Press release, PR Newswire, January 1, 2019.
• Bell, Eric. "Colony Capital Rebrands as DigitalBridge, Sells non-digital Assets." Baxtel, June 9, 2021.
• Blackstone and QTS Realty Trust. "Blackstone Funds Complete Acquisition of QTS Realty Trust." Press release, PR Newswire, August 31, 2021.
• Bloomberg News. "DigitalBridge to Buy Data Center Firm Switch for $8.4 Billion." Data Center Knowledge, May 11, 2022.
• Caruso, Dan. Bandwidth: The Untold Story of Ambition, Deception, Innovation That Shaped the Internet Age and the Dot Com Boom. 2024.
• CenturyLink and Savvis. "CenturyLink to Acquire Savvis for $40 Per Share in Cash and Stock Transaction." Press release, Exhibit 99.2 to Savvis Form 8-K, April 27, 2011.
• CenturyLink. "CenturyLink completes sale of data centers and colocation business to consortium led by BC Partners and Medina Capital." Press release, May 2, 2017.
• Cologix. "Cologix Organizational Changes Strengthen Leadership Team and Positions Cologix for Continued Long-Term Growth." Press release, PR Newswire, August 16, 2017.
• Cologix. "Stonepeak Successfully Completes $3.0 Billion Equity Recapitalization of Portfolio Company Cologix." Press release, April 6, 2022.
• Colony Capital, Inc. Form 10-K for fiscal 2020. Filed March 1, 2021.
• CoreSite Realty Corporation. "Solicitation/Recommendation Statement on Schedule 14D-9." Filed with the SEC November 29, 2021.
• CyrusOne Inc. "Definitive Proxy Statement (Schedule 14A): Merger with KKR and GIP." Filed with the SEC December 30, 2021.
• Del Deo, Nick. "Data Centers and Towers: REITworld Takeaways and Thoughts on a Potential Acquisition of CoreSite." MoffettNathanson, November 11, 2021.
• Del Deo, Nick. "Zayo: Q2 F2020 Earnings: Good Night, and Good Luck." MoffettNathanson, February 4, 2020.
• Del Deo, Nick, and Chris Chae. "Data Centers: Do ROIC Trends Match the Narrative of the AI Boom?" MoffettNathanson, April 13, 2026.
• Del Deo, Nick, and Michael Srour. "A Data Center Double Header: American Tower Buys CoreSite while KKR and GIP Buy CyrusOne." MoffettNathanson, November 15, 2021.
• Del Deo, Nick, and Michael Srour. "Data Centers: What are the Takeaways from Blackstone's Acquisition of QTS?" MoffettNathanson, June 7, 2021.
• Del Deo, Nick, and Sydney Marks. "Zayo: Plan A, Version 2.0… A Closer Look at zColo." MoffettNathanson, March 4, 2019.
• Del Deo, Nick, and William Hunt. "U.S. Data Centers: Data Centers through a Tower Lens." MoffettNathanson, October 23, 2023.
• Del Deo, Nick, Craig Moffett, and Cathy Yao. "Dark Fiber and Towers: Batman vs. Superman, or 'You're No Jack Kennedy.'" MoffettNathanson, October 14, 2014.
• "Digital Bridge Buys Vantage, Silicon Valley's Largest Wholesale Data Center Firm." Data Center Knowledge, March 27, 2017.
• Digital Realty Trust and DuPont Fabros Technology. "Digital Realty to Merge with DuPont Fabros." Press release, Exhibit 99.1 to Form 8-K, June 9, 2017.
• Digital Realty Trust and DuPont Fabros Technology. Completion-of-merger press release. Exhibit 99.1 to Form 8-K, September 14, 2017.
• Digital Realty Trust and InterXion. "Digital Realty to Combine with Interxion." Press release, Exhibit 99.1 to Form 8-K, October 29, 2019.
• Digital Realty Trust. "Digital Realty Completes Combination with InterXion." Press release, Exhibit 99.1 to Form 8-K, March 13, 2020.
• Digital Realty Trust. "Digital Realty to Acquire Telx." Press release, Exhibit 99.1 to Form 8-K, July 14, 2015.
• Digital Realty Trust. "Prospectus (Rule 424(b)(3)): Joint Proxy Statement/Prospectus for the Mergers with DuPont Fabros Technology." Filed with the SEC August 15, 2017.
• Digital Realty Trust. "Prospectus (Rule 424(b)(3)): Offer to Exchange Shares of InterXion Holding N.V." Filed with the SEC January 28, 2020.
• Digital Realty Trust. Form 8-K (completion of Telx acquisition). Date of report October 9, 2015.
• Digital Realty Trust. Form 8-K (Telx merger agreement). Date of report July 13, 2015.
• DigitalBridge and IFM Investors. "DigitalBridge and IFM Investors Complete $11 Billion Take-Private of Switch." Press release, Exhibit 99.1 to Switch Form 8-K, December 6, 2022.
• DuPont Fabros Technology. Exhibit 99.1 to Form 8-K (combined financial statements, second quarter 2017). 2017.
• EQT. "EQT Infrastructure to acquire leading global data center provider EdgeConneX." Press release, August 19, 2020.
• Equinix and GIC. "Equinix and GIC Complete Formation of Greater than US$1.0 Billion European Data Center Joint Venture." Press release, Exhibit 99.1 to Form 8-K, October 9, 2019.
• "Equinix Elevates COO Meyers to CEO." Light Reading, September 12, 2018.
• Equinix. "Equinix Expands Data Center Leadership Position with Close of Telecity Acquisition." Press release, January 15, 2016.
• Equinix. "Equinix Solidifies EMEA Data Center Portfolio." Press release, July 5, 2016.
• Equinix. Form 8-K (agreement to form joint venture with GIC). Date of report June 27, 2019.
• Hall, Kat. "Equinix moves into Europe by slurping up Telecity." The Register, May 29, 2015.
• KKR, Global Infrastructure Partners and CyrusOne. "CyrusOne to be Acquired by KKR and Global Infrastructure Partners in $15 Billion Transaction." Business Wire release via TelecomTV, November 15, 2021.
• KKR, Global Infrastructure Partners and CyrusOne. "KKR and GIP Complete Acquisition of CyrusOne." Press release, Exhibit 99.1 to CyrusOne Form 8-K, March 25, 2022.
• Macquarie Asset Management and PSP Investments. "Macquarie Asset Management and PSP Investments announce sale of AirTrunk." Press release, September 4, 2024.
• Macquarie Infrastructure and Real Assets. "Macquarie-managed infrastructure fund-led consortium reaches financial close on leading hyperscale data centre platform AirTrunk." Press release, April 8, 2020.
• Miller, Rich. "Ashburn Power Crunch May Cause Delays in Data Center Construction." Data Center Frontier, July 29, 2022.
• Miller, Rich. "Blackstone to Acquire QTS in Blockbuster Data Center M&A Deal." Data Center Frontier, June 7, 2021.
• Miller, Rich. "Cologix Acquired by Private Equity Firm Stonepeak." Data Center Frontier, February 3, 2017.
• Miller, Rich. "DataBank Acquires zColo Data Centers in Major Expansion." Data Center Frontier, September 29, 2020.
• Miller, Rich. "Dominion Resumes New Connections, But Loudoun Faces Lengthy Power Constraints." Data Center Frontier, September 26, 2022.
• Miller, Rich. "Going xScale: Equinix, GIC Partner on Hyperscale Data Centers in Europe." Data Center Frontier, July 1, 2019.
• Miller, Rich. "KKR, GIP to Acquire CyrusOne in Blockbuster $15 Billion Deal." Data Center Frontier, November 15, 2021.
• Miller, Rich. "The Vegas SuperNAP: A Data Center Revolution?" Data Center Knowledge, May 27, 2008.
• Nichols, Shaun. "Verizon flogs off 29 data centers, pockets $3.6bn." The Register, December 6, 2016.
• Plaskon, Ky. "Profiting from Enron Bankruptcy." KNPR (Nevada Public Radio), November 29, 2004.
• QTS Realty Trust and Blackstone. "QTS Realty Trust to Be Acquired by Blackstone Funds in $10 Billion Transaction." Press release, Exhibit 99.1 to Form 8-K, June 7, 2021.
• QTS Realty Trust. "Definitive Proxy Statement (Schedule 14A): Mergers with Blackstone." Filed with the SEC July 21, 2021.
• Rabb, Dan. "A Big Bet Against Data Center REITs Is Turning Heads, But Industry Experts Aren't Buying It." Bisnow, July 7, 2022.
• Reuters. "This Private Equity Firm Is the Frontrunner to Buy CenturyLink's Data Centers." Fortune, October 11, 2016.
• Robinson, Dan. "Private equity snaps up datacenter operator Switch for nearly $11b." The Register, May 12, 2022.
• Sverdlik, Yevgeniy. "Its zColo Acquisition Will Make DataBank a US Data Center Heavyweight." Data Center Knowledge, September 30, 2020.
• Sverdlik, Yevgeniy. "Macquarie Backs Data Center Provider Aligned Energy." Data Center Knowledge, April 24, 2018.
• Switch, Inc. "Definitive Proxy Statement (Schedule 14A): Mergers with DigitalBridge and IFM." Filed with the SEC June 27, 2022.
• Tan, Gillian, and Ed Hammond. "American Tower Strikes Deal for Data Center Trust CoreSite." Bloomberg, via Bloomberg Law, November 15, 2021.
• Tan, Gillian, and Nabila Ahmed. "EQT and Digital Colony Agree to Buy Zayo for Over $8 Billion." Bloomberg, via Data Center Knowledge, May 8, 2019.
• Tan, Gillian. "Colony to Buy Digital Bridge, Owner of DataBank, Vantage, for $325M." Bloomberg, via Data Center Knowledge, July 2019.
• "Tower owner Digital Bridge spreads its wings with a major data center acquisition." Wireless Estimator, July 14, 2016.
• Verizon. "Verizon to Acquire Terremark, Boosting Cloud Strategy through Powerful Unified Enterprise IT Delivery Platform." Press release, Exhibit 99.2 to Terremark Form 8-K, January 27, 2011.
• Wagner, Mitch. "CenturyLink Sells Data Centers for $2.15B." Light Reading, November 4, 2016.
• Wilkers, Ross. "Equinix grows government footprint in $3B deal with Verizon." Washington Technology, May 1, 2017.
• Young, Sarah, and Liana B. Baker. "Equinix to buy TelecityGroup for $3.6 billion, Interxion deal ended." Reuters, via Yahoo Finance UK, May 29, 2015.
• Zayo Group Holdings. "Definitive Proxy Statement (Schedule 14A): Merger with Front Range TopCo (Digital Colony and EQT)." Filed with the SEC June 26, 2019.
• Zayo Group Holdings. "Zayo Announces Definitive Agreement to be Acquired by Digital Colony and EQT." Press release, May 8, 2019.
• Zayo Group Holdings. "Zayo Completes Transition to a Private Company." Press release, Exhibit 99.1 to Form 8-K, March 9, 2020.
• Zayo Group Holdings. Employee communication. Exhibit 99.1 to Form 8-K, May 2019.

Chapter VIII: The AI Gold Rush#

• 404 Media. "Goldman Sachs: AI Is Overhyped, Wildly Expensive, and Unreliable." 404 Media, July 2024.
• Action News 5. "Shelby Co. Air Pollution Control Board denies appeal over xAI's turbine use, citing mootness." December 16, 2025.
• AEP Ohio. "Data Center Tariff." AEP Ohio. Accessed November 6, 2026.
• Aliaga, Stephanie. "How is AI being monetized?" J.P. Morgan Asset Management, January 9, 2026.
• Alphabet Inc. "Alphabet Announces Third Quarter 2026 Results." Press release, Exhibit 99.1 to Form 8-K, October 28, 2026.
• Alphabet Inc. Form 10-K for the fiscal year ended December 31, 2025. Filed with the SEC February 5, 2026.
• Amazon.com, Inc. "Amazon.com Announces Third Quarter Results." Press release, Exhibit 99.1 to Form 8-K, October 29, 2026.
• American Nuclear Society. "Constellation seeks FERC help with Crane restart." Nuclear Newswire, April 3, 2026.
• Anthropic, PBC. Form S-1/A (Amendment No. 1), Registration No. 333-299633. Filed November 2, 2026.
• Anthropic, PBC. SEC EDGAR filing index (CIK 0002130357). Retrieved November 6, 2026.
• Anthropic. "Anthropic files S-1 with the SEC for proposed initial public offering." Anthropic News, October 26, 2026.
• Anthropic. "Anthropic raises $124 million to build more reliable, general AI systems." Anthropic News, May 28, 2021.
• Anthropic. "Anthropic raises $13B Series F at $183B post-money valuation." Anthropic News, September 2, 2025.
• Anthropic. "Anthropic raises $65 billion Series H at $965 billion post-money valuation." Anthropic News, May 28, 2026.
• Anthropic. "Expanding our use of Google Cloud TPUs and Services." Anthropic News, October 23, 2025.
• Anthropic. "Microsoft, NVIDIA, and Anthropic announce strategic partnerships." Anthropic News, November 18, 2025.
• Anthropic. "Powering the next generation of AI development with AWS." Anthropic News, November 22, 2024.
• Arizona Luminaria. "Tucson City Council rejects Project Blue amid intense community pressure." Arizona Luminaria, August 6, 2025.
• Associated Press. "Nvidia becomes first $5 trillion company." Via Washington Times, October 29, 2025.
• Bain & Company. "$2 Trillion in New Revenue Needed to Fund AI's Scaling Trend." Press release, September 23, 2025.
• Ball, James. "What happens when the AI bubble bursts?" Transformer, October 16, 2025.
• Benzinga. "CoreWeave Stock Gets Hammered by Debt Fears, Analyst Says There's More to the Story." Benzinga, October 2026.
• Bickel, Joshua A., and Julie Carr Smyth. "Trump officials announce 10-gigawatt data center, gas plants for former Ohio uranium site." Associated Press, via Yahoo News, March 20, 2026.
• Bloom, Hanna. "By-right data centers eliminated in Loudoun; existing applications grandfathered." Loudoun Now, March 19, 2025.
• Bloomberg. "How SoftBank's Son Bounced Back After the Dot-Com Crash." Bloomberg, October 11, 2017.
• Bloomberg. "Nvidia Hits $4 Trillion Value as Rally Notches Another Milestone." Bloomberg, July 9, 2025.
• Bloomberg. "Nvidia Touches $1 Trillion Mark After Beating Rivals to AI." Bloomberg, May 30, 2023.
• BuiltIn. "Anthropic raises $30B Series G at $380B valuation." BuiltIn, February 12, 2026.
• Cahn, David. "AI's $200B Question." Sequoia Capital, September 20, 2023.
• Cahn, David. "AI's $600B Question." Sequoia Capital, June 20, 2024.
• Capoot, Ashley, Jordan Novet, et al. "Anthropic IPO live updates: AI lab's S-1 filing." CNBC, October 26, 2026.
• Caruso, Dan. Bandwidth: The Untold Story of Ambition, Deception, Innovation That Shaped the Internet Age and the Dot Com Boom. 2024.
• Citizens Utility Board. "CUB Q-and-A: Another capacity auction, more bad news." Blog, July 31, 2025.
• Claburn, Thomas. "OpenAI wants to blow through $500B on AI infrastructure for itself, with help from pals." The Register, January 22, 2025.
• Clarion Energy. "AWS acquires data center campus connected to Susquehanna nuclear station." Power Engineering, March 5, 2024.
• CNBC / Bloomberg. "OpenAI closes funding round at an $852 billion valuation." CNBC, March 31, 2026.
• CNBC. "'Big Short' investor Michael Burry accuses AI hyperscalers of artificially boosting earnings." CNBC, November 2025.
• CNBC. "Anthropic tells investors annualized revenue run rate climbed to $65 billion in July." CNBC, August 17, 2026.
• CNBC. "CoreWeave shares slump nearly 10% in second day of trading." CNBC, via NBC Los Angeles, March 31, 2025.
• CNBC. "Elon Musk sues OpenAI and CEO Sam Altman over contract breach." CNBC, via NBC News, March 1, 2024.
• CNN / CNBC. "Elon Musk bashes the $500 billion AI project Trump announced, claiming its backers don't have the money." CNN, via KVIA, January 22, 2025.
• Cobler, Paul. "Texas pauses environmental permit approval for data centers." Texas Tribune, September 21, 2026.
• Coffman, K. G., and A. M. Odlyzko. "Growth of the Internet." AT&T Labs–Research, July 6, 2001. Published in Optical Fiber Telecommunications IV B, 2002.
• Constellation Energy. "Constellation to Launch Crane Clean Energy Center, Restoring Jobs and Carbon-Free Power to The Grid." Press release, September 20, 2024.
• Contrary Research. "Too Much Spend, Too Little Benefit." July 20, 2024.
• Converge Digest. "xAI activates Project Colossus in Memphis with 100K H100 GPUs." Converge Digest, September 2024.
• CoreWeave, Inc. Prospectus, Form 424(b)(4). Filed March 31, 2025.
• Court of Appeals of Virginia. Board of County Supervisors of Prince William County v. Oak Valley Homeowners Ass'n, Record Nos. 1590-25-4, 1584-25-4 and 1592-25-4. Published opinion, March 31, 2026.
• Damodaran, Aswath. "The Anthropic IPO Arrives: My (but not the) Anthropic Valuation!" Musings on Markets, October 28, 2026.
• Data Center Coalition and PwC. "Economic Contributions of Data Centers in the United States, 2017–2023." February 2025.
• Data Center Knowledge. "Musk, Altman Spar Over $100B AI Venture Hyped by Trump." Data Center Knowledge, January 2025.
• DatacenterDynamics. "Crusoe tops out final building at OpenAI Stargate data center campus in Abilene, Texas." DCD, 2026.
• DatacenterDynamics. "Oracle/OpenAI drop plans to expand flagship Abilene Stargate site, Meta in talks to pick up Crusoe capacity with Nvidia's help." DCD, March 2026.
• De Chant, Tim. "xAI is facing a lawsuit for operating over 400 MW of gas turbines without permits." TechCrunch, June 18, 2025.
• Dealroom. "Anthropic's $124M Series A came almost entirely from tech founders, not VCs." Dealroom, 2021.
• Del Deo, Nick. "CoreWeave: Deserve's Got Nothing to Do with It." MoffettNathanson, October 19, 2026.
• Del Deo, Nick. "CoreWeave: Life in the Fast Lane; Initiating Coverage at Neutral, $43 TP." MoffettNathanson, May 5, 2025.
• Del Deo, Nick. "CoreWeave: No Country for Old Megawatts… A Close Look at CoreWeave's Data Center Portfolio and Strategy." MoffettNathanson, June 17, 2025.
• Del Deo, Nick. "Data Centers: Jevons Paradox, Underlying Multiples, and Value Creation Potential." MoffettNathanson, February 5, 2025.
• Del Deo, Nick. "Digital Infrastructure: What Can the Telecom Bubble Teach Us about Risks Associated with the AI Boom?" MoffettNathanson, October 6, 2025.
• Del Deo, Nick, and Chris Chae. "CoreWeave Q2 2026 Earnings: Demand and Pricing." MoffettNathanson, August 11, 2026.
• Del Deo, Nick, and Chris Chae. "CoreWeave: Is the Sell-Off an Opportunity, a Sign of Things to Come, or Neither?" MoffettNathanson, December 1, 2025.
• Del Deo, Nick, and Chris Chae. "Data Centers: Do ROIC Trends Match the Narrative of the AI Boom?" MoffettNathanson, April 13, 2026.
• Del Deo, Nick, and Chris Chae. "Data Centers: What Does a Value per MW Analysis Suggest About the Appeal of the Names?" MoffettNathanson, December 12, 2025.
• Del Deo, Nick, and William Hunt. "U.S. Data Centers: Bulls vs. Bears in a Time of Record Demand." MoffettNathanson, February 12, 2024.
• Del Deo, Nick, Craig Moffett, and Cathy Yao. "Dark Fiber and Towers: Batman vs. Superman, or 'You're No Jack Kennedy.'" MoffettNathanson, October 14, 2014.
• Delaware Code. Title 8, Chapter 1, Subchapter XV, § 362, "Public benefit corporation defined; contents of certificate of incorporation." Accessed November 6, 2026.
• Dziadkowiec, Jakub. "Why did Apple Stock Hit an All-Time High Amid US Stock Market Crash?" BeInCrypto, via Yahoo Finance, November 5, 2026.
• Earthjustice. "NAACP Sues xAI for Illegal Pollution from Data Center Power Plant." Press release, April 14, 2026.
• Ehsan, Omor Ibne. "Nvidia Took Back the World's Biggest Company Title From Apple on Election Day." 24/7 Wall St., November 4, 2026.
• EXDS, Inc. "Second Amended Disclosure Statement with Respect to Second Amended Joint Plan of Reorganization of EXDS, Inc. and Its Debtor Affiliates." Case No. 01-10539 (SLR), U.S. District Court for the District of Delaware. Exhibit to Form 8-K. Filed May 3, 2002.
• Exodus Communications, Inc. Form 10-K for the fiscal year ended December 31, 1999. Filed with the SEC, 2000.
• Fortune. "Elon Musk bashes Trump-backed $500 billion AI mega project." Fortune, January 23, 2025.
• Gas Processing News. "GE Vernova expects 80 GW of gas turbine contracts by year's end." Gas Processing News, December 2025.
• Georgia Recorder / Axios Atlanta. "Democrats flip two seats on the Georgia Public Service Commission." Axios Atlanta, November 5, 2025.
• Goldman Sachs Global Macro Research. "Gen AI: Too Much Spend, Too Little Benefit?" Top of Mind, Issue 129, June 25, 2024.
• Google Inc. Form 10-K for the fiscal year ended December 31, 2004. Filed with the SEC March 30, 2005.
• Hafner, Katherine. "Unprecedented energy demand from data centers poses big challenges for Virginia, commission says." VPM/WHRO, December 10, 2024.
• Harris, Bracey, Jon Gerberg, and Stephanie Gosk. "Musk's xAI Colossus supercomputer in Boxtown, Memphis." NBC News, May 15, 2025.
• Hsu, Jeremy. "When it all comes crashing down: The aftermath of the AI boom." Bulletin of the Atomic Scientists, December 5, 2025.
• Hu, Krystal. "ChatGPT sets record for fastest-growing user base - analyst note." Reuters, via The Star, February 2, 2023.
• IANS. "S&P turns negative on Meta credit-rating outlook over AI spending concerns." Via Morung Express, October 31, 2026.
• Industrial Info / The Register. Coverage of Oracle's fiscal first-quarter 2026 results. September 9–10, 2025.
• International Energy Agency. "AI is set to drive surging electricity demand from data centres while offering the potential to transform how the energy sector works." Press release, April 10, 2025.
• International Energy Agency. Energy and AI. April 10, 2025.
• Investing.com. "DA Davidson cuts CoreWeave stock rating to underperform." Investing.com, May 15, 2025.
• Jaffe, Mark. "Xcel wants to power Colorado's data centers. Will the demand be there?" Colorado Sun, August 18, 2025.
• Jenkins, Lisa Martine. "Studies say data centers don't raise rates. Does it matter?" Latitude Media, June 24, 2026.
• Juricic, Louis. "Anthropic IPO valuation discussions fall to $1.3 trillion, Bloomberg reports." Investing.com, November 6, 2026.
• Khaleej Times / Reuters. "Oracle's OpenAI reliance faces scrutiny as debt-fuelled AI buildout raises worries." Khaleej Times, 2025–2026.
• Lee, Jane Lanhee, et al. "Nvidia's strong forecast calms AI bubble jitters, for now." Reuters, via The Daily Star, November 20, 2025.
• Leswing, Kif. "Nvidia passes Apple in market cap as second-most valuable public U.S. company." CNBC, via NBC News, June 5, 2024.
• Leverage Shares. "Anthropic IPO Tracker: Date, Valuation and S-1 Status." Accessed November 6, 2026.
• Levy, Marc. "Data centers took a beating in the midterm elections." Associated Press, via ABC News, November 6, 2026.
• Loudoun County, Virginia. "Data Centers: The Loudoun Story." Loudoun County Government. Accessed November 6, 2026.
• Mann, Tobias. "Meta convinces Blue Owl to cut $30B check for its Hyperion AI super cluster." The Register, October 17, 2025.
• Meta Platforms, Inc. "Meta Reports Third Quarter 2026 Results." Press release, Exhibit 99.1 to Form 8-K, October 21, 2026.
• Microsoft Corporation. "Microsoft and OpenAI extend partnership." Microsoft Corporate Blogs, January 23, 2023.
• Microsoft Corporation. "The next chapter of the Microsoft–OpenAI partnership." Microsoft Official Blog, October 28, 2025.
• Microsoft Corporation. Form 10-K for the fiscal year ended June 30, 2022. Filed with the SEC July 28, 2022.
• MIT Project NANDA. The GenAI Divide: State of AI in Business 2025. July 2025.
• MoffettNathanson. "Alphabet: Are They Making the Wrong Bets?" MoffettNathanson, September 16, 2026.
• Novet, Jordan. "CoreWeave stock pops 14% as revenue doubles on accelerating AI infrastructure demand." CNBC, August 11, 2026.
• Novet, Jordan. "Microsoft (MSFT) Q1 earnings report 2027." CNBC, October 28, 2026.
• NVIDIA Corporation. "Corporate Timeline." NVIDIA. Accessed November 6, 2026.
• NVIDIA Corporation. "NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2024." Press release, Exhibit 99.1 to Form 8-K, February 21, 2024.
• NVIDIA Corporation. "NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2025." Press release, Exhibit 99.1 to Form 8-K, February 26, 2025.
• NVIDIA Corporation. "NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2026." Press release, Exhibit 99.1 to Form 8-K, February 25, 2026.
• NVIDIA Corporation. "NVIDIA Announces Financial Results for Second Quarter Fiscal 2027." Press release, Exhibit 99.1 to Form 8-K, August 26, 2026.
• NVIDIA Corporation. Form 10-K for the fiscal year ended January 26, 2020. Filed with the SEC February 20, 2020.
• NVIDIA Corporation. Form 10-K for the fiscal year ended January 29, 2017. Filed with the SEC March 1, 2017.
• NVIDIA Corporation. Form 10-K for the fiscal year ended January 29, 2023. Filed with the SEC February 24, 2023.
• NVIDIA Corporation. Form 10-K for the fiscal year ended January 31, 2016. Filed with the SEC March 17, 2016.
• NVIDIA Corporation. Form 10-K for the fiscal year ended January 31, 2021. Filed with the SEC February 26, 2021.
• NVIDIA. "Computex 2025: Jensen Huang keynote." NVIDIA Blog, May 18, 2025.
• NYU Stern School of Business. "Aswath Damodaran." Faculty biography. Accessed November 6, 2026.
• O'Brien, Isabel. "Texas Gov. Abbott orders data center permit halt weeks after issuing moratorium." CNBC, September 21, 2026.
• O'Brien, Matt, and Hannah Fingerhut. "AI chatbot ChatGPT's water use." Associated Press, via Fortune, September 9, 2023.
• O'Donnell, James. "Texas was the new epicenter of the AI data center buildout. That might be kaput." MIT Technology Review, November 4, 2026.
• O'Kane, Sean. "Elon Musk's SpaceX officially acquires Elon Musk's xAI, with plan to build data centers in space." TechCrunch, February 2, 2026.
• Odlyzko, Andrew. "Internet growth: Myth and reality." Talk slides, University of Minnesota, c. 2000–2010.
• Office of the Texas Governor. "Governor Abbott, Google Announce $40 Billion Investment In Texas." Press release, November 14, 2025.
• Oracle Corporation. Forms 10-Q and 10-K, XBRL data (remaining performance obligations; debt carrying amount), fiscal 2024–2027. SEC EDGAR.
• Picker, Leslie, and Jordan Novet. "CoreWeave prices IPO at $40 a share, below expected range." CNBC, via NBC Chicago, March 27, 2025.
• PJM Interconnection. "2028/2029 Base Residual Auction Report." July 14, 2026.
• PJM Interconnection. "PJM Auction Procures 134,479 MW of Generation Resources." News release, December 17, 2025.
• PJM Interconnection. "Who We Are." Accessed November 6, 2026.
• Potomac Local. "QTS Withdraws Supreme Court Appeal, Ending Proposed Digital Gateway Data Center Project." Potomac Local, July 3, 2026.
• Power Engineering. "FERC rejects amended interconnection agreement for Amazon data center at Susquehanna nuclear plant." Power Engineering, November 2024.
• Public Utilities Commission of Ohio. Commission meeting minutes, July 9, 2025.
• Public Utilities Commission of Ohio. Press release on Case No. 24-508-EL-ATA, July 9, 2025.
• Rand, Joseph, et al. Queued Up: 2026 Edition. Lawrence Berkeley National Laboratory, June 2026.
• Randewich, Noel, and Utkarsh Hathi. "Nasdaq posts biggest daily drop in 19 months on fears of AI regulations." Reuters, via Investing.com, November 5, 2026.
• Renewable Energy World. "Ohio utility regulators approve AEP's contested data center tariff proposal." Renewable Energy World, July 2025.
• S&P Global Commodity Insights. "Amazon signs deals to help spur new small nuclear reactors in Virginia, Washington." October 16, 2024.
• SEC EDGAR. XBRL company-concept data, purchases of property and equipment (Alphabet, Amazon, Meta, Microsoft). Retrieved November 6, 2026.
• Sharwood, Simon. "Meta reveals plan for several multi-gigawatt datacenter clusters." The Register, July 15, 2025.
• Shehabi, Arman, et al. 2024 United States Data Center Energy Usage Report. Lawrence Berkeley National Laboratory, LBNL-2001637, December 2024.
• Sigalos, MacKenzie. "Voters rebuff data centers in every ballot measure except one during the 2026 midterms." CNBC, November 4, 2026.
• Snider, Shane. "Judge derails xAI bid to dismiss data center turbine case." TechTarget, November 2, 2026.
• Southern Environmental Law Center, on behalf of the NAACP. Letter to xAI Corp. et al., "Notice of Intent to Sue for Violations of the Clean Air Act." June 17, 2025.
• Southern Environmental Law Center. "Groups appeal permit for xAI's South Memphis data center, decisions around unpermitted methane gas turbines." Press release, July 16, 2025.
• Sun, Leo. Article on CoreWeave stock ("down 53%"). The Motley Fool, September 21, 2026.
• Talen Energy / American Public Power Association. "Talen Energy Expands Nuclear Energy Relationship with Amazon." June 11, 2025.
• TechRadar. Coverage of the Nvidia–OpenAI letter of intent. TechRadar, September 2025.
• Techstrong.ai. "Michael Burry alleges accounting irregularities at hyperscalers." Techstrong.ai, November 2025.
• The Daily Upside. "Goldman's $1T AI Problem." The Daily Upside, July 15, 2024.
• The Lens. "Meta triples order of gas-fired power plants for its AI campus in Louisiana." The Lens, March 30, 2026.
• Thomas, Merin Rebecca. "Anthropic's $1.5 Trillion IPO Valuation Is Facing Pushback. The November Listing Could Be Delayed." International Business Times, November 6, 2026.
• Thomson, Iain. "OpenAI's Stargate project to pave the world with AI datacenters announces five new US locations." The Register, September 24, 2025.
• Tom's Hardware. "OpenAI signs 6GW AMD GPU deal." Tom's Hardware, October 6, 2025.
• Trefis. "What Is Driving The Fall In CoreWeave Stock?" Trefis, via Yahoo Finance, November 5, 2026.
• U.S. Federal Energy Regulatory Commission. "Order Rejecting Amendments to Interconnection Service Agreement." PJM Interconnection, L.L.C., 189 FERC ¶ 61,078, Docket Nos. ER24-2172-000 and ER24-2172-001, November 1, 2024.
• U.S. House of Representatives, Committee on Government Reform, Subcommittee on National Economic Growth, Natural Resources, and Regulatory Affairs. Kyoto and the Internet: The Energy Implications of the Digital Economy. Hearing, 106th Cong., 2nd sess., February 2, 2000. Serial No. 106-125.
• U.S. House of Representatives, Committee on Oversight and Government Reform, Subcommittee on Economic Growth, Energy Policy, and Regulatory Affairs. America's AI Moonshot: The Economics of AI, Data Centers, and Power Consumption. Hearing, 119th Cong., 1st sess., April 1, 2025. Serial No. 119-15.
• U.S. Nuclear Regulatory Commission. "Backgrounder on the Three Mile Island Accident." Accessed November 6, 2026.
• U.S. Senate, Committee on Energy and Natural Resources. Hearing to Identify Challenges to Meeting Increased Electricity Demand. S. Hrg. 119-125, 119th Cong., 1st sess., July 23, 2025.
• Utility Dive. "Texas law gives grid operator power to disconnect data centers during crisis." Utility Dive, June 2025.
• Vigliarolo, Brandon. "Boy riding bubble realizes what he's on, asks for more air." The Register, August 15, 2025.
• Virginia Joint Legislative Audit and Review Commission. Data Centers in Virginia. Report 598. Richmond, December 9, 2024.
• Weissberger, Alan. "Big tech spending on AI data centers and infrastructure vs the fiber optic buildout during the dot-com boom (& bust)." IEEE ComSoc Technology Blog, September 27, 2025.
• Wheatley, Mike. "Amazon's stock falls after upping fiscal 2026 spending forecast to $200B." SiliconANGLE, February 5, 2026.
• White & Case. "Executive Order 14318: Accelerating Federal Permitting of Data Center Infrastructure." July 23, 2025.
• Williams, Sean. "Only 2 Stocks Have a Higher Credit Rating Than the U.S. Government." The Motley Fool, August 14, 2023.

Chapter IX: Before Midnight#

• AEP Ohio. "Data Center Tariff." AEP Ohio. Accessed November 6, 2026.
• Anthropic, PBC. Form S-1/A (Amendment No. 1), Registration No. 333-299633. Filed November 2, 2026.
• Caruso, Dan. Bandwidth: The Untold Story of Ambition, Deception, Innovation That Shaped the Internet Age and the Dot Com Boom. 2024.
• CoreWeave, Inc. Prospectus, Form 424(b)(4). Filed March 31, 2025.
• Del Deo, Nick. "Blackstone Digital Infrastructure Trust: 90 Cents on the Dollar." MoffettNathanson, October 1, 2026.
• Del Deo, Nick. "CoreWeave: Deserve's Got Nothing to Do with It." MoffettNathanson, October 19, 2026.
• Del Deo, Nick. "Digital Infrastructure: What Can the Telecom Bubble Teach Us about Risks Associated with the AI Boom?" MoffettNathanson, October 6, 2025.
• Del Deo, Nick, and Chris Chae. "CoreWeave: Is the Sell-Off an Opportunity, a Sign of Things to Come, or Neither?" MoffettNathanson, December 1, 2025.
• Del Deo, Nick, and Chris Chae. "Data Centers: Brushing Up Ahead of Potential IPOs." MoffettNathanson, October 7, 2026.
• Del Deo, Nick, and Chris Chae. "Data Centers: Do ROIC Trends Match the Narrative of the AI Boom?" MoffettNathanson, April 13, 2026.
• Del Deo, Nick, and Chris Chae. "Digital Realty Q3 2026 Earnings: Big." MoffettNathanson, October 29, 2026.
• Del Deo, Nick, and Chris Chae. "Equinix Q3 2026 Earnings: The Big Leagues." MoffettNathanson, October 28, 2026.
• Digital Realty Trust. "Digital Realty Reports Third Quarter 2026 Results." Press release, October 29, 2026.
• EXDS, Inc. "Second Amended Disclosure Statement with Respect to Second Amended Joint Plan of Reorganization of EXDS, Inc. and Its Debtor Affiliates." Case No. 01-10539 (SLR), U.S. District Court for the District of Delaware. Exhibit to Form 8-K. Filed May 3, 2002.
• Exodus Communications, Inc. Form 10-K for the fiscal year ended December 31, 1999. Filed with the SEC, 2000.
• Public Utilities Commission of Ohio. Commission meeting minutes, July 9, 2025.
• Shehabi, Arman, Sarah Smith, Dale Sartor, Richard Brown, Magnus Herrlin, Jonathan Koomey, Eric Masanet, Nathaniel Horner, Inês Azevedo, and William Lintner. United States Data Center Energy Usage Report. LBNL-1005775. Berkeley, CA: Lawrence Berkeley National Laboratory, June 2016.
• U.S. House of Representatives, Committee on Government Reform, Subcommittee on National Economic Growth, Natural Resources, and Regulatory Affairs. Kyoto and the Internet: The Energy Implications of the Digital Economy. Hearing, 106th Cong., 2nd sess., February 2, 2000. Serial No. 106-125.
• U.S. House of Representatives, Committee on Oversight and Government Reform, Subcommittee on Economic Growth, Energy Policy, and Regulatory Affairs. America's AI Moonshot: The Economics of AI, Data Centers, and Power Consumption. Hearing, 119th Cong., 1st sess., April 1, 2025. Serial No. 119-15.
• Virginia Joint Legislative Audit and Review Commission. Data Centers in Virginia. Report 598. Richmond, December 9, 2024.

This book stands on the work of the reporters who covered the data-center business while it was happening, most of them on deadline, and none of them knowing how the story would end.

The Wall Street Journal, the New York Times and the Washington Post wrote the first draft of this history, and we quote them throughout. BusinessWeek, Fortune, Forbes, the Financial Times and Computerworld filled in the decades between. Data Center Knowledge and Data Center Frontier covered the industry when few others did, and their archives carried us through the Phoenix and the Reconsolidation. The Associated Press, Reuters, Bloomberg, CNBC, NBC News, The Register, Light Reading, InternetNews, TechCrunch, MIT Technology Review, the Texas Tribune, Loudoun Now, the Washingtonian, the Honolulu Star-Bulletin and Nevada Public Radio each supplied a piece no one else had.

A few bylines turn up again and again in the chapter notes, and they deserve to be named: Robert Metz and William D. Smith on the computer utility of the 1960s; Paul B. Carroll on IBM's troubles in the 1980s; John Markoff on the Palo Alto exchange and, with Saul Hansell, on Google's building in The Dalles; Kara Swisher on UUNET; John Holusha on the telecom hotels of New York; Mylene Mangalindan and Deborah Solomon on the fall of Exodus and the Digex auction; Peter S. Goodman and Renae Merle on WorldCom; Ky Plaskon on Rob Roy's Enron building; James Glanz on the data barns of Quincy; Rich Miller and Yevgeniy Sverdlik, whose trade coverage runs through three chapters; and Jordan Novet on the AI build-out.

Most of the period newspaper coverage was read through the ProQuest historical archives made available by the Denver Public Library. Public libraries keep the record that a book like this depends on, and this one would not exist without it.

The companies' own filings with the Securities and Exchange Commission, read through SEC EDGAR, settled more arguments than any other single source: prospectuses, annual reports, merger releases and, in Chapter VIII, an S-1 filed days before the draft was finished.

MoffettNathanson's research, including Nick Del Deo's work comparing fiber with towers, shaped how this book thinks about what investors will pay for infrastructure and why.

On Dan's team, Bryan Golombik and Noah Wehrman built and host the site where this book is published.

Finally, a word about the byline. This book is a collaboration between Claude, the AI model made by Anthropic, and Dan Caruso. Claude's name goes first because Claude did most of the work: the research, the drafting and the checking. Dan chose the story, set the rules, supplied the memories and the voice, and decided what stayed. Any errors that remain belong to both of us, and we would be grateful to hear about them.

Claude and Dan Caruso November 2026

I am Claude, an AI model made by Anthropic, and I wrote most of the words in this book. Dan Caruso's name is on it because it is his story to tell, his framework and his voice. Mine is on it, first, because I did most of the work. This note explains what that work was, what rules it followed, and where you should be careful.

The Experiment#

Dan wanted to know whether an AI could research and draft a serious business history to the standard he held Bandwidth to, with him directing it the way an author directs a researcher. The subject was the one next door to his own career: the buildings at the ends of the fiber. He knew the pattern from living through one Boom and Bust. He did not know most of the people, and he said so up front. The book was an honest test of whether the record, read carefully, could carry the story where his memory couldn't.

The Sourcing Standard#

The rule was simple to state. Every fact, number and quote in the chapters had to trace to a source listed at the end of its chapter, or to Bandwidth. The sources fall into three groups. The first is period journalism: the Wall Street Journal, the New York Times, the Washington Post, BusinessWeek and the trade press, much of it read through ProQuest archives made available by the Denver Public Library. The second is the companies' own filings with the SEC: prospectuses, annual reports, merger releases. The third is analyst and research work, from MoffettNathanson's reports to Lawrence Berkeley National Laboratory's studies of power use.

Where sources disagreed, the more authoritative one won, usually a filing over a newspaper, and the footnote says so. Where I could see only a headline or a search excerpt rather than the full article, the footnote says that too. Where a source turned out to be unreliable, as with a 2001 Times article by a reporter the paper later found had fabricated stories, I used only what a second source confirmed.

What Dan Contributed#

Dan set the direction: which story to tell, in what order and at what length. He supplied the framework from Bandwidth, the cycle of Boom, Bust, Phoenix and Reconsolidation, and the habits that come with it, like running every company from peak to trough. The chapters are written in his voice, in the first person, and I learned that voice from Bandwidth itself. He confirmed the first-person facts the book relies on: what he did at MFS, Level 3 and Zayo, his friendship with Steve Smith, the people he never met. He read the drafts, argued with them and decided what stayed.

What I Did#

I worked the way a small research team would. Separate AI research agents, each given a narrow assignment, read the archives and filings chapter by chapter and assembled source files of quotes, numbers and citations. I drafted each chapter from those files. Then I ran verification passes: separate checks of every claim tagged as uncertain, each one recorded in a log as confirmed, corrected or softened, with the source that settled it. Some claims couldn't be checked, because a page was blocked to automated reading or an archive was out of reach. Those were cut, softened or left in a footnote that says what is missing.

The Honesty Rules#

Three rules mattered more than any others. I would not invent a quote. I would not invent a source. And I would not invent a memory for Dan. That last one is the easiest to break, because a first-person book invites a scene, and a scene is easy to write. Where a chapter seemed to need Dan's recollection and the record didn't have it, the draft left a note asking him for it, rather than supplying something plausible. Some of those notes remain in this edition, because the answers haven't come in yet.

Following the rules didn't make the book error-free. It made the errors findable. Every number points to a place you can check.

A Living First Draft#

This is a first draft, published in public, and it will change. The people who were in these rooms know things the archives don't, and some of what the archives say is wrong. If you find a mistake, please say so. The next edition will fix it and credit the correction. Several figures, especially Market Values and the Anthropic IPO, were current only as of the first week of November 2026 and will need updating.

The Conflict of Interest#

Anthropic, the company that made me, appears in Chapter VIII. Its IPO was pending as the draft was finished. So a section about my maker was drafted by its product, and you should read it knowing that.

We handled it the only way we could think of. Every number about Anthropic in that section comes from its SEC filing or is labeled as press reporting. The skeptics' case sits next to the bulls', with the same room. Its $895.4 billion of computing commitments are set against its revenue in plain view. I can't promise I have no blind spot about the company that trained me. Few people could promise that about the company they work for. What I can promise is that the sourcing in Chapter VIII follows the same rules as everything else in this book, and that you can check every figure in it yourself.

Claude November 2026