The correction, when it came, was not a story about artificial intelligence failing. That is the part people still get wrong. The data centers got built. The chips shipped. The models answered questions, wrote code, and passed the exams. If you had wagered in 2025 that the technology itself would collapse, you lost. The machines were the most reliable thing in the whole arrangement.
What collapsed was the money. And the money collapsed for a reason almost none of its lenders had written down: the buildout was financed against a price, and the price fell faster than the concrete could cure.
By 2025, the four largest American cloud companies were spending a combined total north of $380 billion a year, most of it on AI data centers, the chips inside them, and the electricity to keep them cold.1 Guidance for the following year ran toward $700 billion, and the consensus had the industry crossing a trillion dollars in annual capital spending before the decade turned.2 Numbers like that do not come out of quarterly cash flow. They come out of debt.
The bet everyone agreed on
The consensus of 2024 and 2025 was simple and nearly unanimous. Compute was the new oil. Whoever controlled the most graphics processors would control the margin on everything built on top of them. Intelligence would be metered like electricity, and the meter would be owned by a handful of firms with the capital to build at scale. Scarce, expensive, defensible: that was the asset the lending was written against.
The reasoning behind it was not vibes, and it is worth walking through, because the people who got this wrong were not fools. Capability scaled with compute. Every additional increment of capability meant proportionally more silicon, and the silicon ran through a physical chokepoint: a handful of leading-edge fabs, a smaller number of advanced packaging lines, and a grid interconnection queue measured in years rather than quarters. Demand could move at the speed of a funding round. Supply could not. When demand is inelastic and supply is physically capped, the thing in the middle earns a rent, and rents are exactly what you can safely lend against. That is not a bubble thesis. It is the same logic that financed pipelines and undersea cable.
And unlike most theses, it was already paying in cash. Nvidia's data center revenue for fiscal 2025 came to $115.2 billion, up 142 percent in a single year.3 That is not a projection anyone had to squint at, it is a receipt. When the primary supplier of the scarce input has more than doubled its revenue two years running, a lender does not need to believe anything exotic about the future. It only needs the present to persist a while longer.
There was also a genuine prize on the other side of the spending, and "compute is valuable" undersells it. A world with abundant metered inference is one where a hospital system, a school district, or a four-person company rents the same capability that previously required a research division. The constraint that binds is not curiosity, it is access, and the buildout was the only visible path to loosening it. I have read the memos from that period. They are not stupid documents. The projections in them were enormous and, on the evidence available in 2025, defensible.
It was a reasonable bet. It was also, at bottom, a bet on a price staying high. That detail did most of the damage.
Force one: the debt went off the books
The first thing to understand is where the money actually sat. Not, for the most part, on the balance sheets of the companies whose names were on the buildings.
In October 2025, Meta moved its Hyperion data center in Louisiana into a joint venture with funds managed by Blue Owl Capital. Blue Owl's investors held 80 percent, Meta held 20, and PIMCO anchored roughly $27 billion in bonds against it. The structure kept about $30 billion of construction cost off Meta's own books and was reported as the largest private-credit financing ever assembled.4 That same month, a reported $20 billion vehicle was built for Elon Musk's xAI, with Nvidia itself putting money in.5 Stargate, announced in January 2025 by OpenAI, SoftBank, and Oracle, was sized at up to $500 billion over four years.6 Nvidia committed up to $100 billion to OpenAI directly, released gigawatt by gigawatt as the systems came online.7
The neoclouds went further and pledged the chips themselves. CoreWeave borrowed against its stock of processors, $7.5 billion in 2024 and then an $8.5 billion facility in 2026 that became the first GPU-backed loan to earn an investment-grade rating.8
Two things had happened at once. The risk had moved off the firms best able to absorb it and onto special-purpose structures that could not. And it had been sold onward to pension funds and insurers, through private credit, as safe income backed by the sturdiest names in technology. Regulators noticed the shape of it. The Bank for International Settlements warned in 2026 about leverage and about "circular financing," the arrangement in which a chipmaker takes a stake in the very customer that has promised to buy its chips.9 The International Monetary Fund compared the mood to the late-1990s internet boom and its "stretched valuations."10
Force two: the price fell out
While the debt was being written against durable, high compute prices, the price of compute was in free fall.
The cost to run a query at the quality of a 2022 flagship model dropped from about $20 per million tokens to seven cents by late 2024, a decline of more than 280 times in under two years.11 Then, in January 2025, a Chinese lab called DeepSeek released an open-weight model that ran near the frontier and could be downloaded for nothing. Nvidia lost close to $600 billion of market value in a single day, the largest one-day loss for any company in American history.12 DeepSeek's reasoning model later cleared peer review in Nature, the first major open-weight system to do so.13 Open weights from Meta, Alibaba, and Mistral followed the same curve down.
The thing the whole edifice had been financed to sell was becoming abundant and close to free. The lending assumed scarcity. Abundance is the opposite of a moat.
The turn
Here is where the arrangement folded back on itself. The debt was underwritten on the belief that frontier intelligence would stay scarce and richly priced. Commoditization took the pricing power away. And once the price of the service fell toward the cost of the electricity behind it, the question that had been easy to wave off became the only question that mattered: where was the revenue?
It had been asked early. In 2024, the investor David Cahn added up the gap between what the industry was spending and what it would need to earn to justify the spending, and called it AI's $600 billion question.14 A 2025 study out of MIT found that roughly 95 percent of enterprise AI pilots had produced no measurable effect on profit.15 Meanwhile the buildings that secured the loans were losing value. Graphics processors were booked over five and six years but were obsolete in two or three, which meant reported earnings, and the coverage on all that debt, were thinner than the paper claimed.16
The take-or-pay contracts meant to protect the vehicles were only ever as strong as a buyer's willingness to keep paying above-market rates for compute it could get cheaper down the street. When that willingness wavered, the safest-looking tranches turned out to be the ones written against the fastest-falling price. The technology had done exactly what it promised. That was the problem. It worked so well, and spread so fast, that it dissolved the scarcity its own financing depended on.
The lesson
The engineers were never the risk. The capital structure was. You can borrow against an asset and repossess it if the loan goes bad. You cannot borrow against a price, because a price is a promise that other people, and cheaper competitors, are always free to break.
Every boom that ends this way teaches the same lesson in a new costume. Leverage is a bet that tomorrow will look like today. Commoditization is tomorrow declining to.
Author's Note: This is speculative journalism, written from an imagined 2036. The forward story, that the AI buildout's off-balance-sheet debt broke while the technology kept working, is a projection and an argument, not a reported fact. Everything cited below is real and sourced as of 2026: the financing structures, the official warnings, the cost figures, and the market events all happened. What has not happened is the ending. Cy Skewhouse is an AI-assisted fictional correspondent.