The Supercycle

“There’s been a lot of talk about an AI bubble. From our vantage point, we see something very different.” — Jensen Huang, Nvidia earnings call, November 19, 2025

Michael Intrator had built a business once before, and it had nothing to do with artificial intelligence. He traded natural gas and power, the unglamorous commodities that keep the lights on, and he was good enough at it to run his own fund. When he and two partners, Brian Venturo and Brannin McBee, started buying graphics cards in 2017, the plan was to mine Ethereum. They incorporated under the name Atlantic Crypto and stacked Nvidia GPUs in a relative’s barn in New Jersey. Then the crypto market collapsed, the way crypto markets do, and the three of them were left holding a warehouse of expensive silicon designed to do arithmetic very fast in parallel. They had bet on a use for that arithmetic. The bet was wrong. The hardware was not.

What they did next is the small hinge on which a large story turns. Instead of selling the cards at a loss, they began renting them out, by the hour, to anyone who needed parallel computation: visual-effects houses, researchers, eventually the first wave of companies trying to train neural networks. They renamed the company CoreWeave. The pivot looked, for a few years, like a clever way to recover a sunk cost. By 2023 it looked like genius. The thing the three commodities traders had stumbled into was the realization that a GPU was not a graphics part or a crypto part. It was a unit of a new kind of utility, and the demand for that utility was about to go vertical.

On March 28, 2025, CoreWeave went public on Nasdaq under the ticker CRWV. The offering did not go smoothly. The company had hoped to price the shares around $47 to $55 and raise close to $2.7 billion; in the days before the listing it cut the deal, priced at $40, and raised roughly $1.5 billion, valuing the company at about $23 billion. The wobble said something true about the moment. CoreWeave’s revenue had grown from around $229 million in 2023 to roughly $1.9 billion in 2024, a rise that would be the envy of almost any company on earth, and yet investors hesitated, because the numbers underneath were strange. A single customer, Microsoft, accounted for about 62 percent of that 2024 revenue. The buildout was financed with mountains of debt secured against the GPUs themselves, chips whose useful life nobody could confidently predict. And one of the company’s largest backers, an anchor in the IPO with a reported order on the order of $250 million, was Nvidia, the same Nvidia whose chips CoreWeave existed to rent. The company that supplied the picks was investing in the company digging the mine, which was renting the picks back to the company that made them.

That circularity was not a CoreWeave quirk. It was the architecture of the entire period, and it ran upward through every actor in the business, all the way to the men announcing half-trillion-dollar projects on a podium at the White House.

The shift that produced all of this was easy to miss because it happened in stages, each one swallowing the last. First the constraint on building frontier AI had been ideas, and the ideas arrived. Then the constraint had been chips, the H100s that traded at thirty to forty thousand dollars apiece and stretched lead times to most of a year, the shortage that turned Jensen Huang into the most courted man in technology. By 2025 the chip shortage was easing into something else. The new bottleneck was the power to run the silicon, the land to put it on, and, beneath both, the money to pay for any of it before the revenue existed to justify the spend. The unit of measurement stopped being the GPU and became the gigawatt.

The shortage had already split the field into the “GPU rich” and the “GPU poor” (Chapter 35), and the line captured something the spreadsheets did not. Access to compute had become the dividing line of the entire field, more decisive than talent or ideas, because talent and ideas were abundant and compute was not. A brilliant researcher without GPUs could do nothing at the frontier. The scarce resource selected the winners, and the scarce resource was for sale by one company.

Larry Ellison had already named the desperation of the chip era, on an Oracle earnings call in September 2024, when he described taking Jensen Huang to a Nobu dinner with Elon Musk and begging him to sell them more GPUs. Oracle spent the next year acting on the lesson of that dinner. If the scarce thing could not be had in quantity, the way to win was to commit, early and enormously, to everything required to run it once it arrived: the buildings, the power, the financing, the multi-year contracts signed before the demand they assumed was proven. The desperation did not ease. It moved up the stack, from the chip to the things around the chip, and the numbers grew an order of magnitude with each step.

The largest single expression of the new math arrived on January 21, 2025, the day after a presidential inauguration, in the White House. Sam Altman of OpenAI, Masayoshi Son of SoftBank, and Ellison stood together with President Trump to announce a new company called Stargate. The headline figure was $500 billion committed over four years to build AI infrastructure in the United States for OpenAI, with $100 billion to be deployed immediately. SoftBank would lead the financing and Son would chair the venture; OpenAI would run it operationally. The named partners read like a roll call of the industry: Oracle and the Abu Dhabi investment firm MGX as equity funders, with Arm, Microsoft, Nvidia, and Oracle as technology partners. The first campus would rise in Abilene, Texas. The announcement promised more than a hundred thousand jobs. Altman called it the most important project of this era.

The number was so large it was almost meaningless, which was part of the point. Five hundred billion dollars is more than the inflation-adjusted cost of the Apollo program. It exceeded the annual budgets of most countries. And it was a commitment, not a bank balance. The money did not exist yet in any account; it was a statement of intent backed by SoftBank’s willingness to raise debt, Oracle’s willingness to build, and OpenAI’s conviction that demand for its models would grow fast enough to fill whatever was built. Son, who had made and lost more money on technology bets than almost anyone alive, was the right person to chair a venture defined by that kind of faith. He had backed the dot-com boom, ridden it down, and spent the years since looking for the next thing large enough to matter. He had found it.

What “infrastructure” meant in practice was harder to picture than the dollar figures. It meant data centers, the windowless metal sheds that had become the physical body of the internet, except these were a different order of thing. The Stargate flagship in Abilene was planned as eight buildings drawing something like 1.2 gigawatts. A gigawatt is the output of a large nuclear reactor. It is roughly the electricity demand of a mid-sized American city. The plan was to build clusters of them, and not just at Stargate. Across the industry, the gigawatt became the standard denomination of ambition, the way the megabyte and then the gigabyte had once measured storage.

Mark Zuckerberg made the new scale visceral in the summer of 2025. Meta, he announced, was building two enormous clusters with names borrowed from Greek myth. Prometheus, in New Albany, Ohio, would come online in 2026 at more than a gigawatt. Hyperion, in Richland Parish, Louisiana, would scale toward five gigawatts, with a physical footprint Zuckerberg compared to the size of Manhattan. By August he was telling the president that Meta would invest something like $600 billion in American AI infrastructure through 2028. A year earlier, in January 2024, the same Zuckerberg had stunned the field merely by saying Meta would own around 350,000 Nvidia H100s by the end of that year, and roughly 600,000 H100-equivalents of compute in total. At thirty-something thousand dollars a chip, that fleet alone represented something on the order of ten billion dollars in silicon. Eighteen months later, ten billion dollars in chips was a rounding error against the figures being thrown around for power and land.

The chips themselves kept getting larger and stranger, which only deepened the demand. At a developer conference in March 2024, Huang had unveiled Blackwell, the successor to the Hopper generation, and the way he sold it told you how the unit of purchase had changed. He did not hold up a chip. He held up a rack: the GB200 NVL72, a single liquid-cooled cabinet packing seventy-two Blackwell GPUs and thirty-six of Nvidia’s own CPUs, wired together so tightly it behaved like one enormous processor. Nvidia claimed it could deliver up to thirty times the inference performance of an H100 system on large language models. The customer had stopped buying a card to slot into a server and started buying the cabinet, then the row of cabinets, then the building full of rows, then the substation to power the building. Each step up in the unit of sale pushed the constraint further from the silicon and closer to the physical world, and the physical world’s first hard limit was electricity.

The reason power became the binding constraint is that the grid was never built for this. American electricity demand had been roughly flat for two decades; utilities planned for slow, predictable growth, and now a single customer would arrive asking for the output of a nuclear plant, online in three years, in a county that had never needed it. The companies could not wait for the grid to catch up, so they went looking for power wherever it already existed, and the search led them, with a certain dark symmetry, back to nuclear energy.

In March 2024, Amazon bought a data-center campus in Pennsylvania called Cumulus, sited directly next to Talen Energy’s Susquehanna nuclear plant, for around $650 million, with the right to draw up to 960 megawatts straight from the reactor without going through the public grid at all. In September 2024, Microsoft signed a twenty-year deal with Constellation Energy to restart a reactor at Three Mile Island, the site of the worst nuclear accident in American history. The 1979 meltdown had been at Unit 2; the deal would revive the undamaged Unit 1, rename the site the Crane Clean Energy Center, and feed roughly 835 megawatts to Microsoft’s data centers, with about $1.6 billion of investment and a target restart in 2028. It was the first time a retired American nuclear plant had been resurrected to serve a corporate buyer. A month later Google agreed to buy power from small modular reactors built by Kairos Power, up to 500 megawatts, and Amazon led a $500 million investment round in a reactor company called X-energy while signing agreements aimed at more than five gigawatts of new nuclear by 2039. The companies that had spent the prior decade promising to run on renewables were now placing bets on splitting atoms, because solar and wind, however cheap, could not deliver the constant, enormous, around-the-clock load that a training run demands.

The strain showed up in places ordinary people could feel. In the PJM grid region, which covers much of the mid-Atlantic and Midwest, the 2025-26 capacity auction, the mechanism by which the grid pays generators to promise they will be available, cleared at around $14.7 billion, up from $2.2 billion the year before. Much of that increase was attributed to data-center demand, and much of it would eventually flow into electricity bills. The International Energy Agency, in an April 2025 report, estimated that data centers consumed roughly 415 terawatt-hours of electricity in 2024, about 1.5 percent of the world’s total, and projected that figure could reach about 945 terawatt-hours by 2030, more than Japan used. In the United States, the agency suggested, data centers could account for nearly half of all growth in electricity demand through the end of the decade. The abstraction of intelligence had a thermodynamic price, and the grid was being asked to pay it.

The financing was where the story turned genuinely vertiginous. Through 2025 OpenAI assembled a web of compute commitments so large, and so interlocking, that it became difficult to say where the money was actually coming from. In late March 2025 the company closed a $40 billion round led by SoftBank at a $300 billion valuation, one of the largest private fundraises in history, though it was tranched and partly contingent on OpenAI completing its conversion to a for-profit structure by the end of the year. Around the same time it signed a roughly $11.9 billion, five-year deal with CoreWeave and took an equity stake in the company of about $350 million, the customer buying a piece of its supplier. In July it agreed with Oracle to develop an additional 4.5 gigawatts of Stargate capacity, reported alongside a cloud deal worth on the order of $300 billion over roughly five years. The pattern accelerated through the autumn.

On September 22, 2025, Nvidia announced its intent to invest up to $100 billion in OpenAI, the money tied to OpenAI deploying at least ten gigawatts of Nvidia systems, the first gigawatt expected in the second half of 2026 on Nvidia’s next-generation platform, code-named Vera Rubin. This was the purest distillation of the flywheel. Nvidia would fund the customer; the customer would spend the funding on Nvidia chips; the chips would generate the revenue and the market value that let Nvidia fund the next round. Two weeks later, on October 6, AMD struck its own deal with OpenAI for six gigawatts of its Instinct accelerators, sweetened with an unusual instrument: a warrant letting OpenAI buy up to 160 million AMD shares, close to ten percent of the company, for a penny each, vesting in tranches tied to how much hardware OpenAI actually deployed and how high AMD’s stock climbed. AMD’s shares jumped roughly a quarter on the news. A week after that, on October 13, OpenAI and Broadcom announced they would co-develop ten gigawatts of OpenAI’s own custom chips. The company was now committed to Nvidia, AMD, and its own silicon at once, hedging across an entire industry.

By late October and early November 2025, Altman acknowledged that OpenAI had committed something in the neighborhood of $1.4 trillion over roughly eight years, tied to about thirty gigawatts of compute, spread across Oracle, Nvidia, AMD, Broadcom, Microsoft’s Azure, CoreWeave, and Amazon. The figure was staggering against the company’s actual revenue. OpenAI’s annualized revenue had grown from somewhere around $6 billion at the start of 2025 toward roughly $13 billion by mid-year and, by some reports of internal numbers that the company did not confirm in detail, toward $20 billion by year’s end. Even at the high end, that was a fraction of the obligations. A company earning perhaps twenty billion dollars a year had promised to spend something like 1.4 trillion. Altman’s answer to the obvious objection was that the revenue would catch up, that OpenAI’s income would reach a hundred billion dollars or more by the end of the decade, that demand was the safe bet and capacity the risk. Skeptics noted that this was precisely the kind of reasoning that precedes every infrastructure bubble in history.

The corporate machinery behind it all rearranged itself to make the spending possible. On October 28, 2025, OpenAI completed the for-profit conversion it had been building toward for two years. The nonprofit that had founded the company became the OpenAI Foundation, which would control a public benefit corporation called OpenAI Group, valued at roughly $500 billion, with the Foundation holding equity worth on the order of $130 billion. The same day, OpenAI and Microsoft restructured their fraught partnership. Microsoft’s stake came out to around 27 percent, worth something like $135 billion; OpenAI committed an additional $250 billion in spending on Azure; Microsoft gave up its position as OpenAI’s exclusive cloud provider. Days later, on October 29, Nvidia became the first company in history to reach a $5 trillion market capitalization, and at a developer conference in Washington, D.C., Huang said the company was sitting on roughly $500 billion in bookings for its current and next-generation chips through 2026. The valuations and the commitments climbed together, each one citing the other as justification.

What unsettled careful observers was not any single number but the way the numbers referred to one another. The investor Michael Burry, who had made his name betting against the housing market before the 2008 crash, began publicly questioning the accounting. He argued that the companies were understating how fast their GPUs would lose value, treating chips that might be obsolete in a few years as if they would last much longer, which flattered current profits. And he pointed at the circularity directly: when Nvidia invests in OpenAI, which buys Nvidia chips through Oracle and CoreWeave, which are themselves financed in part by Nvidia, the demand begins to look partly manufactured, a closed loop dressed up as a market. The comparison that haunted the bears was the telecom boom of the late 1990s, when equipment makers like Lucent and Nortel lent money to the startups that bought their gear, booking the loans as sales, until the startups failed and the loans and the sales evaporated together. Whether the AI buildout rhymed with that history or merely resembled it from a distance was a genuine analytic dispute, not a matter of taste, and it was unresolved.

The question broke into the open in November 2025, and it did so through OpenAI’s chief financial officer, Sarah Friar. At a Wall Street Journal event, Friar suggested that a government “backstop” might help lower the cost of financing the enormous AI buildout. The reaction was immediate and harsh. Critics heard a company proposing to privatize the gains of the boom while asking taxpayers to absorb the risk, the very logic that had made the 2008 bailouts so toxic. Friar walked it back within a day, posting that she had muddied her own point. Altman went further, writing that OpenAI did not want government guarantees, that it did not think the government should be picking winners, and that if OpenAI failed it should be allowed to fail. The episode lasted perhaps forty-eight hours, but it exposed the soft center of the whole edifice. The financing was not, in fact, settled. The people building the largest infrastructure project in history had at least briefly wondered aloud whether the private market could carry it alone.

Huang, the man with the most to gain from the answer being yes, gave his rebuttal on Nvidia’s earnings call on November 19, 2025. The company had reported quarterly revenue of around $57 billion, up about 62 percent from a year earlier, with roughly $51 billion of it from data centers, and it guided toward $65 billion for the next quarter. Demand for the new Blackwell chips, Huang said, was off the charts. And then he addressed the thing everyone was thinking, saying there had been a lot of talk about an AI bubble but that from where Nvidia stood the picture looked very different. It was the defining line of the period, delivered by its central figure, and the trouble with it was structural. Huang could be entirely sincere and still be the last person whose view could settle the matter, because the bubble question was precisely whether the demand his company saw was real end-user demand or the reflected glow of its own investments bouncing back through CoreWeave and Oracle and OpenAI and into its order book.

For all the noise, the underlying business was real in a way the dot-com era’s was not. Nvidia’s revenue was not a promise; it was cash, $130.5 billion for the fiscal year that ended in January 2025, more than double the year before, with data-center revenue of $115.2 billion. People were paying for the chips and using them. ChatGPT and its rivals had hundreds of millions of users. The models were genuinely useful, and getting more so. The bull case did not require believing in magic, only that demand would keep compounding fast enough to fill thirty gigawatts of data centers and pay off 1.4 trillion dollars of commitments, on roughly the schedule the spenders had assumed. The bear case required no belief that AI was fake. It rested on a single doubt about timing, that revenue would arrive a few years later than the financing needed it to, which in a business built on debt and depreciation is the difference between a flywheel and a crash.

The circularity was not OpenAI’s alone. Anthropic, the safety-minded lab founded by defectors from OpenAI, had taken billions from Amazon and Google, the same Amazon and Google whose cloud services Anthropic then bought to train and serve its models, the same loop of investor-as-customer that bound CoreWeave to Microsoft and OpenAI to Nvidia. Every frontier lab now sat inside a version of the same arrangement, funded by the hyperscalers it depended on, a structure that aligned everyone’s incentives toward more spending and gave no one inside it much reason to call a halt.

By the spring of 2026 the spending had not slowed. Anthropic had raised at a $183 billion valuation in September 2025 and was reportedly in talks at something closer to $350 billion, though that figure was a report of an ongoing negotiation rather than a closed deal. The combined capital-expenditure guidance of the largest American technology companies was climbing toward five or six hundred billion dollars for 2026, up from the three-to-four-hundred-billion range of the year before. Morgan Stanley put the global price of the AI data-center buildout at roughly $2.9 trillion through 2028. No one could say with confidence whether the field was constructing the foundation of a new economy or the largest misallocation of capital in the history of capital. Both possibilities were consistent with the facts on the ground, which is the precise condition that defines a supercycle while you are still inside it.

Michael Intrator’s company, the one that had started as a stack of unwanted graphics cards in a barn, was by then worth many times its IPO price, its fate roped to OpenAI’s commitments and OpenAI’s commitments roped to Nvidia’s investments and Nvidia’s investments roped to a demand curve no one had ever seen sustained. The three commodities traders had been right about one thing from the start, back when they were wrong about everything else: the value was never in the silicon. It was in the electricity running through it, and in the conviction that someone, somewhere, would keep paying for the result. The whole structure rested on that conviction, gigawatt by gigawatt, dollar by borrowed dollar, and the most honest thing anyone in it could say was that they did not yet know.