In May we asked what it costs a frontier AI lab to serve a million tokens, the unit these companies sell their models by. The answer we got was $154, and almost all of it was fixed cost: training runs, researchers, and above all the data centers. Fixed costs get cheaper per token as volume grows, so the question became one of demand. The labs' spending only made sense if the industry roughly doubled the tokens it served. If demand kept compounding, all three frontier labs (OpenAI, Anthropic and Google) would come out fine; if it stalled, one of them probably wouldn't.

Five months later we can check that work against real numbers. Anthropic has filed to go public, and its S-1, the registration statement a company files with the SEC before an IPO, is the first time one of these labs has opened its books. What it shows is better for the labs than our May model suggested: the leading lab now roughly covers all its costs, training included, from what customers pay it, and its contracts need much less growth than it's getting. The strain we were looking for is real, but on our rough numbers it has moved to the companies that build the data centers the labs rent.

Was $154 a Million Tokens Right?

It was too high, mostly because we spread the fixed costs over too few tokens. The May model assumed the whole industry served about 250 trillion output tokens a year. The article itself cited OpenAI's roughly $2 billion of 2024 inference spending, and at the model's own cost per token that implies about 305 trillion for OpenAI alone. Fix just that and $154 becomes about $33 per million tokens.¹ Even that is generous: Google alone now runs about 17 quadrillion tokens a year through its APIs.²

We also had the prices wrong. We said frontier tokens sold for $15 to $75 per million, but on the day we published Anthropic's flagship listed at $25 per million output tokens, and the mid-tier models most developers use now cost $10 to $20.³

And the "1.9x" demand figure, the doubling at the center of the May piece, turned out not to measure demand at all. Token volume cancels out of the formula; what's left is the labs' spending commitments divided by one lab's annual costs.

Some of the May article held up. OpenAI looked like the most exposed of the three labs, and it still does. Free users and heavy users on flat-rate plans looked like the place the losses concentrate, and they still are.²³

Does a Lab Cover Its Costs Yet?

Anthropic roughly does now, training included. Serving customers is very profitable: in the first half of 2026 it cost $2.4 billion against $16.3 billion of revenue. But serving is the small part. Once you add training, research, salaries and sales, Anthropic lost $693 million on that revenue, an operating margin of about −4%, or a profit of about 10% on the company's adjusted measure, which leaves out stock compensation.⁴ A year earlier it lost $2.1 billion on $1.2 billion of revenue, and its cash from operations has since turned positive.

Anthropic's profit will look worse on paper after the IPO. Most of the stock it has granted employees only counts as an expense once the company goes public, and the S-1 says that will add about $23.7 billion of costs, mostly over the next two years.⁴ That's around $11 billion a year, about a sixth of current revenue, enough to put the company back into an operating loss. None of it is cash, so it doesn't change whether Anthropic can pay its compute bill, but it is a real cost to shareholders.

Most of that revenue, 80%, comes from customers paying per token, so falling token prices hit it directly. About half of all Anthropic's revenue reaches it through Amazon, Google and Microsoft, whose cloud customers can buy Anthropic's models through their existing cloud accounts. Those three companies are also Anthropic's investors and the suppliers of its compute, which makes the revenue look circular. The money in those sales still comes from the end customers who buy through those accounts.⁴

OpenAI is further from covering its costs. It spends far more on research and training relative to its revenue: $19.2 billion of R&D in 2025 against $13.1 billion of revenue.²² Its free users add to the gap. Only about 6% of ChatGPT's weekly users pay,⁵ and in 2025 nearly half of OpenAI's $8.4 billion inference bill went to people who don't pay at all.⁶ In the first quarter of 2026, investors' money still covered about 40% of its cash spending, and in the second quarter it lost $12.3 billion on $6.7 billion of revenue.⁷

The labs are responding by moving heavy users onto the meter. GitHub switched its Copilot coding assistant to token-based billing in June, Anthropic now charges business seats for every token used, and OpenAI has reportedly brought back usage caps on its $20 plan.⁸

What Growth Does Anthropic Need to Pay Its Compute Bill?

Anthropic needs its revenue to grow a little under 30% a year, far less than it's growing now.

Anthropic doesn't own most of its computers. It rents them from Google, Amazon, Microsoft and others under long contracts, and most of those contracts require it to pay whether or not it uses the capacity. Those firm commitments come to nearly $400 billion, or about $59 billion a year once they're all running.⁹

With Anthropic's other costs, it needs about $111 billion of revenue in 2028 to cover everything. It earns about $67 billion a year now, so it needs a bit under 30% growth a year to get there.¹⁰ Its own plan, about $195 billion in 2028, takes about 70%.¹¹

It has signed up for another $200 billion of computing on top of that. Anthropic can cancel the largest piece, about $120 billion with SpaceXAI, on 90 days' notice. The rest, with Nscale and Lambda, it can't cancel but pays for only as the capacity is delivered; counting those as firm too raises the growth it needs to about 36% a year.⁹

What Could Go Wrong?

The price of a token could fall faster than the number of tokens grows.

Revenue is tokens sold times the price per token, and the price keeps falling. Ramp, a corporate-card company that tracks what businesses spend on AI, says the price businesses paid per token across the labs fell 41% between its March peak and early September.¹⁴ List prices for the models most developers use fell even faster, by 35% to 60% between May and October.¹² Kept up for a year, Ramp's pace is about a 65% decline, and at that rate Anthropic needs to sell nearly 4 times as many tokens each year just to cover its commitments.

That's about as fast as token volume has been growing. Google's API tokens grew a little more than fourfold in the past year, and its tokens across all products grew 6.7 times in the year to May.²¹³ Those counts include free use, such as the AI answers in Google Search, so they overstate paid demand. Revenue is a better test: Anthropic's grew 45% from the second quarter to the third, a pace of more than four times a year. So far that's enough, but it leaves little room.

So far volume has kept ahead. Anthropic's third-quarter revenue still rose 45% while that price fell. But three months is a short window, and there are early signs of slowing: Anthropic's third quarter was the first since at least early 2025 in which revenue grew by a smaller dollar amount than the quarter before.⁴¹⁰ Anthropic's compute spending, meanwhile, rises as its contracted capacity arrives, whether or not demand has. In the third quarter, compute spending grew about 77% over the second while revenue grew 45%, so the roughly break-even first half may not repeat as more committed capacity comes online.⁴¹⁰

OpenAI faces a harder version of the same test. It needs revenue to grow about 68% a year just to cover its current costs by 2028, and about 90% a year to reach its own plan of $350 billion in 2030.¹⁵ Most of its revenue comes from flat-rate subscriptions, and growing those means signing up more paying users, which cheaper tokens don't do.¹⁶

How Is Google Doing on Both Sides?

Renting out capacity pays Google, but building it now costs more than the company earns. Google is both a frontier lab and one of the landlords the labs rent from, and the May article ended by asking whether selling its chips and data centers to a rival lab was smart or a slow erosion of its advantage. Its third-quarter results give a partial answer.

Google Cloud's revenue doubled to $30.7 billion in the third quarter, and its operating margin rose from 23.7% to 36.2% in a year.² Anthropic is a big part of that: its commitment of at least $114.5 billion is about a fifth of Google Cloud's backlog of signed future business, and Google is also one of Anthropic's shareholders.⁴¹⁷

Google's capital spending more than doubled from a year earlier, to $51.4 billion in the quarter, while cash from operations fell to $42.3 billion. That's the second quarter in a row it spent more than it brought in. To keep building, Google raised $29 billion of debt and sold $12.4 billion of new stock in the quarter, and it stopped buying back its own shares.²

So Where Is the Risk Now?

The risk is now with the companies building the data centers, which is where many critics have been looking all along. Michael Burry, the investor known from The Big Short, has argued that these companies write their chips off over too many years, which flatters their profits.¹⁸

Amazon, Alphabet, Microsoft, Meta, Oracle and others are spending roughly $800 billion a year on AI data centers.¹⁹ For what's been built since 2023 to earn a normal return, customers would need to pay about $430 billion a year, or about $780 billion once the labs and products on top are counted. Our rough estimate is that they pay about $220 billion today.²⁰

Nobody pays that gap directly. The builders absorb it as a lower return on what they've built, and they borrow or issue stock to keep going. Oracle burned $5.4 billion of cash last quarter, and Google is now borrowing and issuing stock to keep up.²¹ If customer spending doesn't grow into all that capacity, the labs would feel it first, since their rent is fixed, and Anthropic's cash would cushion the blow for a while. If the shortfall lasted, much of the loss would reach the builders, who are also the labs' landlords, lenders and shareholders.

What Would Change the Answer?

• Whether the price businesses pay per token keeps falling at Ramp's pace through the fourth quarter.
• Whether Anthropic's growth keeps slowing.
• Whether $125 billion to $150 billion of additional chips Anthropic expects from 2028 to 2031 arrives on schedule. At full run it would add $31 billion to $37 billion a year to the bill.⁴
• Whether Anthropic's cash flow stays positive once it has used up compute it paid for in earlier years.

For readers of the May piece, smaller fixes: Oracle's OpenAI backlog is about 66% of its total, not 54%; the R&D layer was $36 in the text and $34 in the chart; capex was 81% of the full economic cost, not 40%; the "two of three labs survive" finding was never modeled; and about three-quarters of Google's estimated cost advantage came from an assumption about its cost of capital that we never supported, with the TPU itself accounting for less than a tenth.

The model behind this analysis is available as a spreadsheet with editable inputs.

Sources