Chapter IX: Before Midnight

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.