The US AI Labs, Mapped: Founders, Revenue and the Money Behind Them

Source: https://codingplans.dev/briefs/us-model-labs-part-1/

Home>Briefs>The US Model Labs, Part 1 Market Structure Brief · Part 1 of 3 • By codingplans.dev • Dataset verified 2026-10-28

Nine American AI companies carry $2.18T of valuation on $22.2B of 2025 revenue. Investors close that gap in two ways: they pay a steady multiple for revenue that is growing fast, and where there is no revenue they pay for the founder. Part of the compute behind the growth is financed by the labs' own suppliers.

Combined valuation

$2.18T

Latest sourced marks; xAI at its SpaceX merger price

2025 revenue

$22.2B

OpenAI $13.07B, Anthropic $4.59B, xAI ~$3.8B; undisclosed figures excluded

Going rate

~53x

Cognition's valuation over run-rate, in May and again in September

Anthropic

21x

Run-rate multiple at the Series H; 210x on 2025 revenue

On October 26 Anthropic filed its registration statement with the SEC ahead of a Nasdaq listing, and for the first time the price of the most valuable private AI company could be checked against audited accounts. Five months earlier the company had raised $65 billion at a valuation of $965 billion. Its revenue for the whole of 2025 was $4.59 billion.

That is 210 times revenue. Measured the other way the same round allows, it is 21.

The other way is run-rate, recent sales annualised, which Anthropic put above $47 billion when it announced the Series H. The filing shows the run-rate was no one-month spike. Revenue in the first half of 2026 was $16.29 billion, 14 times the same half a year earlier and more than three times all of 2025. When two honest multiples for one company differ tenfold, last year's accounts are not what anyone is paying for. The US labs are priced on the slope of the revenue line, and on a bet about how long it stays that steep.

This part asks what rate the market pays for a slope, and whose money pays it. The nine companies are the American labs whose main business is a model or a product built on one; the AI divisions of Alphabet, Meta and Microsoft do not report revenue separately and are left out. Every figure, with its basis, is in the table at the end.

Fifty times revenue is the going price

The cleanest measurement comes from Cognition, which makes the coding agent Devin, because it was priced twice in four months. In May it raised at $26 billion on a run-rate it put at $492 million. In September it raised again at $48 billion, with the run-rate at about $900 million. Both numbers had nearly doubled. The ratio between them stayed at 53.

The September investors paid the same rate on a bigger number. You do that when you expect the curve to keep its shape.

Most of the market sits near that rate. Runway does not publish revenue, but Sacra estimated it at about $96 million annualised at the end of 2025, which puts its February round at about 55 times. OpenAI's March round valued it at 65 times its 2025 revenue, the audited figure that leaked in June. Because that revenue is a full past year rather than a current rate, the multiple on OpenAI's sales today is lower by however much it has grown since. Perplexity, at 100 times reported recurring revenue in September 2025, is the most expensive of the group.

xAI also comes out at 65 times, on Sacra's estimate of about $3.8 billion annualised. The price behind that ratio, though, was set in February when SpaceX bought xAI in an all-stock deal, and Elon Musk controls both companies. A price agreed between two companies with the same owner is not a market price.

Filled markers use revenue the company disclosed, audited or as a stated run-rate; hollow markers use outside estimates or press reports. Safe Superintelligence and Thinking Machines have no revenue to plot, and Midjourney has no valuation.

On this chart Anthropic is the outlier in both directions. Against its 2025 revenue it is the most expensive company in the market by a wide margin. Against its run-rate it is the cheapest, at less than half the going rate, and it looks cheap for the simple reason that its revenue grew faster than its price.

That is the strongest case for the round, and its whole risk. A multiple of 21 is cheap only for as long as the line keeps climbing.

Without revenue, investors price the résumé

Two of the nine cannot be plotted at all. Safe Superintelligence has released nothing since Ilya Sutskever founded it in 2024, and was last valued at $32 billion in April 2025. In July 2026 NVIDIA announced a "substantial" investment in it and early access to its Vera Rubin systems. Reuters, citing a source, put the amount at $5 billion.

Sutskever explained the deal in one sentence: "We have research that is worthy of scaling up, and having access to a big NVIDIA computer will let us do so."

Thinking Machines Lab is the same bet at a smaller size. Mira Murati founded it in 2025 and raised a $2 billion seed round at $12 billion that July, led by Andreessen Horowitz, before it had a product on the market.

In both cases the asset being bought is a person, and the two people came from the same place. Sutskever co-founded OpenAI and was its chief scientist; Murati was its chief technology officer until 2024. Dario and Daniela Amodei left OpenAI with a group of colleagues to start Anthropic in 2021, and Musk, who co-founded OpenAI in 2015, started xAI in 2023. Three of the nine labs were founded by people who left OpenAI, and a fourth by one of its founders.

The concentration is the mechanism, and it follows from what is scarce. Models are published, distilled and undercut within months; the practical knowledge of how to make a frontier training run work is not. It is held by a small number of people, many of whom learned it at the same company in the same few years. If you cannot hire that team, you fund it, and a founder who ran research at the lab that set the pace gives investors something to underwrite before there is revenue. It is a bet on a slope that has not started.

Midjourney runs the other way. David Holz started it in 2021 and has never taken outside money. It has been profitable since shortly after it launched in 2022, and outside estimates put its 2025 revenue between $300 million and $500 million, which would put it among the mid-sized labs on revenue. It has no valuation, because a valuation is the price the last investor paid and there has been no investor.

A valuation measures a transaction, not a business, and Midjourney is the awkward row for everyone else in the table. It shows that generative AI can pay for itself when it is sold as a finished product to people who pay every month. The labs mostly sell something else, raw capability to developers and companies at frontier cost, and only one of them has begun to show that this pays too. Until the rest do, someone has to pay for the compute.

Suppliers help pay for the slope

Some names recur across the rounds for ordinary reasons. Andreessen Horowitz co-led OpenAI's March round, led Thinking Machines' seed and co-led Cognition's September round, which is a venture firm concentrating its bets. NVIDIA's position is different. It put $30 billion into OpenAI, joined xAI's Series E and invested in SSI, and it sells hardware to all three.

Anthropic's filing shows how far the arrangement goes. Amazon holds convertible notes in the company, bought in 2023 and 2024, that turn into shares at the listing. The S-1 also commits Anthropic to spend at least $110 billion with Amazon through April 2036, and about $114.5 billion with Google, another early investor, through July 2033, whatever its usage. A rival is on the list too: Anthropic has agreed to spend up to $120.5 billion on NVIDIA-based capacity from SpaceX's xAI through 2029, cancellable at 90 days' notice once the first periods end in early 2027.

The cloud providers sit on the revenue side as well. In the first half of 2026, 48 percent of Anthropic's revenue came through the AWS, Google Cloud and Azure marketplaces. Compute and infrastructure cost it $7.33 billion in 2025, more than it earned all year, and the S-1 says part of the build-out "has been funded or supported by capital or credit support provided or facilitated by certain of our CSPs and chip partners."

So the suppliers finance the compute that produces the growth, collect part of the revenue it brings in, and hold equity whose value depends on both.

They do not set every price. Anthropic's $965 billion was led by Altimeter, Dragoneer, Greenoaks and Sequoia, none of which sells it anything, and their judgment counts. But those funds were pricing a revenue line that supplier capital had helped build. The slope is real; part of what made it so steep was paid for by the companies that sell the labs their compute.

The circle can close, and Anthropic's first half is the evidence. Its operating loss in 2025 was $8.06 billion. In the first six months of 2026, on $16.29 billion of revenue, it fell to $693 million, and on the company's adjusted measure, which leaves out stock-based compensation, non-cash charitable contributions and the Bartz copyright settlement, it became an operating profit of $1.62 billion. The much larger headline net loss is mostly an accounting charge on convertible instruments that rises with the company's own valuation. Dependence on a few buyers fell too. The September draft had two customers supplying nearly a quarter of 2025 revenue; the S-1 says no customer accounted for more than 10 percent in the first nine months of 2026.

OpenAI has not shown that turn. Its operating loss was $20.92 billion in 2025, more than double the year before, on revenue that more than tripled.

The listing is where the slope gets tested

Suppliers financing their customers has a precedent with the same mechanism. In the late 1990s Lucent and Nortel lent heavily to the telecom carriers buying their equipment, and the loans went bad when the carriers' revenue did not arrive in time. The comparison holds for the flow of money. It breaks on the revenue, which in Anthropic's case is arriving: the first half of 2026 brought in more than three times the whole of 2025. Where it could still hold is the cost. Anthropic says its compute commitments run to 2045 and generally have to be paid whatever its revenue or utilisation. The S-1 leaves the total blank; the September draft put it at $518 billion or more.

Anthropic has proposed a listing under the ticker ANT, with the price range still to be filled in, and reporting in September put its target above $2 trillion. OpenAI is preparing its own. Once the shares trade, the price is set every day by a market rather than negotiated once in a round.

Two results would change this reading. If the second half of 2026 shows Anthropic's growth slowing sharply, for instance because token prices fall faster than usage grows, 21 times run-rate stops being cheap, and every lab priced at fifty is exposed to the same arithmetic. If the stock settles well below the private mark while revenue is still climbing at this pace, the private rounds were carrying a premium, and the companies that paid for it are the ones on the other side of the compute contracts.

Either way, the suppliers are paid first. The compute bills fall due whatever the slope does; only the shares have to wait to find out what it was worth.

| Company & founders | CEO | Founded | 2025 revenue | Latest valuation | Latest round | Val / revenue | Sources | | --- | --- | --- | --- | --- | --- | --- | --- |

**Part 2 follows the money.** How much of the labs' funding comes from their own suppliers, and what happened the last time that pattern ran at this scale.

Compare the paid plans →

Sources: OpenAI round CNBC, Bloomberg and 2025 accounts Ars Technica, operating loss Techstrong; Anthropic Series H Anthropic and 2025 and 2026 accounts SEC S-1, September leak TNW; xAI Series E, SpaceX merger and Sacra; SSI NVIDIA and Reuters; Thinking Machines Reuters; Perplexity Reuters; Cognition TechCrunch; Midjourney revenue estimates; Runway TechCrunch and Sacra. Checked 2026-10-26.

Got a question, or spotted an outdated price or wrong number? Email hello@codingplans.dev. Every figure links to its source.