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#
| Company | 2023 | 2024 | 2025 | 2026 |
|---|---|---|---|---|
| Alphabet | $32.3B | $52.5B | $91.4B | $195–205B guided ($132.0B through Sept.) |
| Amazon | $52.7B | $83.0B | $131.8B | ~$220B guided ($175.7B through Sept.) |
| Meta | $27.0B | $37.3B | $69.7B | $135–145B guided, incl. finance leases ($83.1B cash through Sept.) |
| Microsoft (fiscal years to June) | $28.1B | $44.5B | $64.6B | $116.0B (FY2026); ~$170B guided for calendar 2027, incl. finance leases |
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.