Customer revenue covers a sliver of the AI build-out, even on the bulls' numbers

Matteo’s essay arguing that the AI bubble may not exist is better than most bear cases are bad, and I went in expecting to disagree for the usual reasons. The usual reasons turned out to be unnecessary. The debate has no definition of a pop, so both sides can read any quarter as confirmation. I will supply one: a pop is when customer revenue stops covering the cost of the capital that built the capacity, and outside investors stop making up the gap. That is a ratio, and the essay hands over enough numbers to compute it.

The author puts committed spending at 3 trillion dollars and spending already done at 1.9 trillion, around 5 trillion in all. Spread over 20 years, that is 250 billion dollars a year. He admits the sum is rough, and it is generous. It ignores interest, and it treats the whole pile as if it were a building. A lot of a data centre budget is chips, and I doubt the chips will be in service when a 20-year loan is paid off.

Now take his best witness. He cites an Anthropic S-1 (I could not check the filing, so these are his figures) showing 16.3 billion dollars of revenue in the first half of 2026, up from 4.585 billion in all of 2025 and 41 million in 2023. Double the half-year to annualise it. The same filing reportedly shows 2.08 dollars of revenue per dollar of compute, so compute takes less than half of each dollar, and what remains has to pay for salaries, sales and everything else. Against 250 billion a year, that remainder is a small fraction. Even on raw revenue, before compute comes out, the figure is small. This is the fastest-growing company in the sample, picked by the author as his proof.

The objection is obvious and he makes it: the bill belongs to the whole industry, and growth at that speed does not stop soon. Fair, so be generous. He imagines AI consumption at ten times today’s level. Give Anthropic alone that multiple on its annualised revenue and the total clears 250 billion, but only before compute is paid for. After compute it still falls short. Covering the whole bill from this one company at that margin needs revenue well beyond even the tenfold case. Other labs will carry part of the load, and OpenAI is the one the author concedes is unknown. Each lab also needs its own margin, and the margin has to come from customers who are not other AI companies.

That last condition is where I think the essay’s second argument fails. He points to Alphabet, where Google Cloud grew 15.5 billion dollars (36%) against 37.8 billion (12%) for Google Services, and concludes that data centre rental does not explain Alphabet’s earnings. That may hold for Alphabet, but Cloud revenue is a supplier’s income and a lab’s cost, and the author himself says OpenAI buys an enormous amount of compute from other companies. Supplier growth shows money being spent, and says nothing about whether anyone downstream earns it back. The first quarter in which labs cut purchases is the first quarter suppliers shrink.

The same goes for his P/E argument. Alphabet, Meta and NVIDIA sit at stable multiples in the 20s, against hundreds for the dot-com names. I take the point about that era. But a price-to-earnings ratio holds only while the earnings do, and if part of those earnings is other AI firms spending borrowed or invested money, the ratio looks cheapest just before the denominator falls.

None of this says Anthropic is a bad business. Revenue growing from 4.585 billion in a year to 16.3 billion in a half-year is remarkable, if the filing says what he says. It does say that a 5 trillion dollar build-out has a long way to travel before customers pay for it.

So run the sum yourself on the next half-year filing: annualise revenue, multiply by one minus 1/2.08, divide by 250 billion. Today it comes out as a small fraction, and until it gets close to one I would not describe the debt as covered by customers, whatever the P/E ratios say.

Prompted by I'm Not Convinced The AI Bubble Exists, Matteo's Website.