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

In May 2026 Anthropic raised $65 billion at a valuation of $965 billion. Its revenue for the whole of 2025, according to the registration statement it filed with the SEC on October 26 ahead of a Nasdaq listing, was $4.59 billion. Put one over the other and the round valued Anthropic at about 210 dollars for every dollar it had earned the year before.

That ratio makes the round look more reckless than it was. In the same announcement, Anthropic said its run-rate revenue, recent sales annualised, had passed $47 billion, which brings the price down to about 21 times. The filing shows that the run-rate was not a one-month spike. Revenue in the first half of 2026 was $16.29 billion, about 14 times the $1.16 billion of the same period a year earlier and more than three times the whole of 2025. Both sets of numbers describe the same company a few months apart, and the distance between them is what this series is about. Last year's accounts tell an investor very little about an AI lab. What the money is buying is the slope of the revenue line, and a bet on how long it stays that steep.

Not every lab has a slope to show. Of the nine companies below, three have large and fast-growing revenue, three sell products built on models and report smaller run-rates, two have no revenue at all, and one has never sold a share to anyone. The list leaves out the AI divisions of Alphabet, Meta and Microsoft, which do not report their revenue separately.

The three with revenue

OpenAI and Anthropic are the only companies here with audited numbers. Anthropic's arrived with its S-1; OpenAI's became public by accident. Financial statements leaked in June and checked by the Financial Times show revenue of $13.07 billion in 2025, up from $3.7 billion the year before. In March the company closed a $122 billion round at $852 billion, about 65 times that revenue. Most of the money came from three companies: Amazon put in $50 billion, NVIDIA and SoftBank $30 billion each. About $35 billion of Amazon's share, Bloomberg reported, only arrives if OpenAI goes public or reaches a technical milestone.

Growth at that pace has been expensive. OpenAI's operating loss was $20.92 billion in 2025, against $8.78 billion a year earlier. Its net loss was bigger still, about $38.5 billion, but most of the difference is a non-cash accounting charge from its conversion to a for-profit company. Anthropic's 2025 accounts had the same shape at a smaller size. Its operating loss grew to $8.06 billion from $2.98 billion. Of its $41.97 billion net loss, $34.31 billion was a change in the fair value of convertible notes and similar instruments, a charge that grows whenever the company's valuation does. Compute and infrastructure cost it $7.33 billion of $12.65 billion in operating costs, more than the company earned all year.

The first half of 2026 looks different. On $16.29 billion of revenue, Anthropic's operating loss shrank to $693 million. On its own adjusted measure, which leaves out stock-based compensation, non-cash charitable contributions and the Bartz copyright settlement, it made an operating profit of $1.62 billion. The net loss for the half was $80.75 billion, but $72.39 billion of that was the same fair-value charge, which the filing ties to the jump in valuation at the Series H. At the end of June the company held $131.65 billion in cash and short-term investments.

The filing also speaks to two risks that a run-rate hides. One is concentration. A leaked draft reported by Reuters in September had nearly a quarter of 2025 revenue coming from two customers. For 2026 the S-1 says no customer accounted for more than 10 percent of revenue in the first nine months, and that the share coming from outside the top five rose from about 64 percent in the first quarter to about 79 percent in the third. The other is cost. Anthropic says its compute commitments run to 2045 and generally have to be paid whatever its revenue or utilisation, but the S-1 leaves the total blank for now. The September draft put it at $518 billion or more.

xAI is the hardest of the three to price. Sacra, a research firm, estimates its annualised revenue at about $3.8 billion at the end of 2025; the company has not confirmed a figure. In January 2026 it raised $20 billion in a Series E. A month later SpaceX, which Elon Musk also controls, bought it in an all-stock deal that valued xAI at $250 billion. That is the number in the table, but it was agreed between two Musk companies rather than by an outside investor, and it should be read that way.

The application companies

The next three sell products built on models: Cognition's coding agent Devin, Perplexity's answer engine and Runway's video tools. Their revenue is smaller and mostly reported as run-rates, but the clearest view of how investors price this market comes from one of them.

In May 2026 Cognition raised money 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 in four months, and the ratio between them barely moved: about 53 times in May, about 53 times in September. The second set of investors paid the same multiple as the first, on a bigger number. That is a reasonable thing to do only if you expect the growth to continue at the same rate.

The other two sit in a similar range. Perplexity was valued at $20 billion in September 2025, on reported annual recurring revenue of about $200 million, or 100 times. Runway does not publish revenue. Sacra estimated it at about $96 million annualised at the end of 2025, a few weeks before Runway raised $315 million at $5.3 billion. If the estimate is close, Runway was priced much like Cognition.

The two with nothing to sell yet

Safe Superintelligence has not released a product since Ilya Sutskever founded it in 2024. It was last valued at $32 billion in April 2025. In July 2026 NVIDIA announced a "substantial" investment and a long-term partnership that gives SSI access to its next generation of Vera Rubin systems. NVIDIA did not say how much; Reuters, citing a source, put it at $5 billion. Sutskever's explanation ran to 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, which Mira Murati founded in 2025, raised a $2 billion seed round led by Andreessen Horowitz at $12 billion that July, before it had anything on the market. With no revenue to multiply, both valuations rest on what the founders know and who they can hire.

Midjourney is the reverse case: real revenue and no valuation at all. David Holz started it in 2021 and has never taken outside money. It has been profitable since shortly after launching in 2022, selling image generation to consumers for $10 to $120 a month, and outside estimates of its 2025 revenue range from $300 million to $500 million. A valuation is the price at which the last investor bought in, and nobody has bought into Midjourney, so it has none. That is why it is missing from the chart below.

Plotted by founding year, the valuations have little to do with age or with revenue. SSI, two years old and without a product, is valued at about 6 times Runway, which was founded in 2018 and sells to large companies. To see why investors accept that, it helps to look at where the founders came from.

Where the founders learned it

For three of the nine, the answer is OpenAI. Dario and Daniela Amodei left with a group of colleagues and started Anthropic in 2021. Sutskever, an OpenAI co-founder and its chief scientist, started SSI in 2024. Murati, OpenAI's chief technology officer until 2024, started Thinking Machines the following year. A fourth line runs through Elon Musk, who co-founded OpenAI in 2015 and started xAI in 2023.

Training a frontier model takes compute, data and a kind of practical knowledge that is still held by a small number of people, many of whom learned it at the same company in the same few years. A founder who ran research at the lab that set the pace offers investors something to underwrite before there is revenue. That is the bet behind SSI and Thinking Machines. It also helps explain why the labs tend to converge on similar methods and read each other's published work closely: a good part of the field trained together.

Who wrote the cheques

Read row by row, the investor lists look varied. Read across, the same names keep appearing. Andreessen Horowitz co-led OpenAI's March round, led Thinking Machines' seed and co-led Cognition's September round. NVIDIA put $30 billion into OpenAI, joined xAI's Series E and invested in SSI.

NVIDIA also sells hardware to all three, and the same arrangement runs through the cloud providers. Amazon committed $50 billion to OpenAI. It also holds convertible notes in Anthropic, bought in 2023 and 2024, that convert into shares as part of the listing, and according to the September leak it is due to sell Anthropic about $110 billion of computing capacity over seven to ten years. Google, another early Anthropic investor, is down for about $111 billion. The cloud companies are Anthropic's sales channel too: 48 percent of its revenue in the first half of 2026 came through the AWS, Google Cloud and Azure marketplaces. Even xAI appears on Anthropic's list of suppliers, with up to $84.5 billion of NVIDIA-based capacity through 2029, most of it cancellable at 90 days' notice. The S-1 itself acknowledges that part of Anthropic's compute build-out "has been funded or supported by capital or credit support provided or facilitated by certain of our CSPs and chip partners." A large part of the money the labs raise goes back to companies that invested in them, as payment for compute.

Suppliers financing their own customers has a history. In the late 1990s Lucent and Nortel lent heavily to the telecom carriers buying their equipment, and many of those loans went bad when the carriers' revenue did not arrive in time. The more immediate test is the stock market. Anthropic's S-1 proposes a Nasdaq listing under the ticker ANT, with the price range still to be filled in; reporting in September put its target above $2 trillion. OpenAI is preparing a listing of its own. Public investors will be the first large buyers of these shares who are not also selling the labs something.