Here is a sentence I had to read three times before I believed it. In 2025, OpenAI booked roughly 13 billion dollars of revenue and posted an operating loss of about 21 billion. The loss was larger than the revenue. Not larger than the profit — larger than the entire top line.
I have written here before about what AI costs the companies that buy it. This is the other half of the ledger: what it costs the companies that sell it, and what happens to everyone else if those numbers do not converge.
The state of the ledger
As of this autumn, the three frontier labs are running three completely different experiments.
OpenAI is the one everyone watches. Its annualised run rate reached about 40 billion dollars by August, up from a long plateau near 25. It also burned roughly 3.7 billion in the first quarter against 5.7 billion of revenue — about sixty-five cents of every dollar. It has around 900 million weekly users and somewhere near five percent of them pay, which is a beautiful consumer story and a brutal margin story, because the free ninety-five percent still cost real money to serve. Internal projections reportedly put cumulative losses near 44 billion through 2028 with profitability no earlier than 2029.
Anthropic is running the opposite experiment and, for now, winning the argument. Second-quarter revenue came in around 11 billion with a first operating profit of roughly half a billion. Its gross margin went from deeply negative in 2024 to somewhere near sixty percent, driven mostly by inference getting cheaper rather than prices going up. Around four fifths of revenue is enterprise, which is stickier and cheaper to serve than consumer.
xAI is the cautionary column. By the most widely cited estimate it loses something like twenty-six dollars for every dollar it takes in.
Three companies, one technology, wildly different unit economics. That alone should end the lazy version of this debate. There is no single AI business model yet. There are experiments with very different survival odds.
The part that is not on the balance sheet
The labs are not where the real money is, though. The real money is in the buildout underneath them, and that is where the accounting gets interesting.
Hyperscaler capital spending is somewhere around 600 billion dollars this year, with some estimates running closer to 800. Whatever the exact figure, it has roughly tripled since 2024. Amazon's capital spending over the last twelve months exceeded its own operating cash flow, which is a polite way of saying free cash flow went negative. Amazon's capex this year is larger than the combined capex of the entire publicly traded US energy sector.
Three details make this more fragile than the headline numbers suggest.
First, a lot of the commitment is off the balance sheet. Moody's put data-centre leases that are signed but not yet commenced at roughly 662 billion dollars — obligations that do not show up in capex figures because of when GAAP starts the clock. That off-book number is larger than the combined on-book debt of the same companies. Independent estimates of total off-balance-sheet AI commitments run to 1.65 trillion, and the Bank for International Settlements has warned that some of the financing structures being used may be obscuring leverage rather than removing it.
Second, depreciation. Hyperscalers write AI hardware down over five or six years. A lot of people who work with these chips think the useful economic life is closer to two or three. If that is right, reported earnings across the industry are currently flattered by something in the order of 176 billion dollars between 2026 and 2028. That is not fraud; it is an assumption. But it is an assumption that unwinds mechanically, on a schedule, whether or not revenue arrives.
Third, circularity. Chipmakers take equity in labs, labs sign take-or-pay compute deals with clouds, clouds buy chips with debt. Money moves in a circle and every loop of it books as somebody's revenue. Some of that demand is real. Some of it is the same dollar being counted several times on its way around the track.
The dot-com comparison gets reached for constantly and it is half right. The scale is bigger. The difference is who is holding the bag: in 2000 the spenders were companies with no profits. Today the spenders are the most profitable companies in history. That does not make a correction impossible. It changes what one would look like.
What running at a loss actually buys
The bull case is Amazon: lose money for years, capture the platform shift, harvest returns forever. It is a good story and it has one weak joint. Amazon's losses bought warehouses and software, assets that kept working and kept earning. AI losses buy compute that is obsolete in thirty-six months and model weights that a competitor matches in six. You have to keep paying to stay in the same place. That is less like building a moat and more like renting one.
The bear case is that the revenue simply never catches the spending, and I think the honest position is that nobody knows, including the people spending it. What we can say is that Anthropic's quarter proved the economics are not inherently broken. Profit at the frontier is possible. It just turned out to require enterprise customers, disciplined inference, and not carrying 850 million free users.
Three ways this goes
Revenue catches up. Enterprise adoption keeps compounding, inference keeps getting cheaper, and the 2027 depreciation wall gets absorbed by growth. The capex looks visionary in hindsight. Entirely plausible.
An orderly correction. Capex growth flattens, some data-centre commitments get quietly renegotiated, there are large write-downs, a few heavily indebted intermediaries fail, and the hyperscalers take a bad year and carry on. This is the most likely version, and it is not a catastrophe. It is a cycle.
A disorderly one. The failure would not start at the labs. It would start at the leveraged middle — the neoclouds and intermediaries carrying tens of billions in debt against depreciating GPUs and contracts with counterparties who are themselves unprofitable. One of those failing is a credit event, not a tech story, and credit events do not stay in their lane.
What to actually watch
Ignore the think-pieces, including this one. There are three checkable signals. Hyperscaler capex growth slowing or reversing for two consecutive quarters. GPU utilisation falling below what current depreciation schedules quietly assume. And a frontier lab taking a serious valuation cut in a funding round instead of a markup. None of those has happened yet. If two of them happen together, the argument is over.
The part that reaches Lagos
Here is why a Nigerian engineer should care about American accounting.
Right now, inference is priced below what it costs, subsidised by investors who are buying market share. Every developer outside the rich world is currently getting a discount funded by venture capital and hyperscaler balance sheets. If the correction comes, the first thing that happens is not that the models disappear. It is that the price moves toward what the compute actually costs. I wrote a few months ago that AI tokens already cost more than a Lagos developer's salary in the markets where those tokens are metered. A repricing makes that worse, quietly, for everyone who was never in the room where the capex was approved.
I keep thinking about the twenty-one against the thirteen. There is nothing inherently wrong with a company losing money to build something; that describes most things worth building. What is new is the ratio, and the fact that the asset being bought wears out faster than the debt used to buy it.
Somebody is going to be right about this and somebody is going to be extremely wrong, and the unusual thing about 2026 is that we will probably find out which within about eighteen months.