OpenAI's Run Rate Is $50B, Not $70B, After Investors Grossed It Up to Match Anthropic

TL;DR

OpenAI told investors its annualized revenue was roughly $50 billion at the end of September, the Financial Times reported on Thursday and CNBC confirmed. That is about $20 billion below the "nearly $70 billion" figure that circulated in late September and got repeated as fact. Nobody lost $20 billion. The bigger number came from OpenAI's own investors grossing up its net revenue so it could be compared with Anthropic, which books cloud-partner sales at full value. The market did not read the footnote: Nvidia fell 3%, Oracle nearly 6% and CoreWeave nearly 8% on the day.

What changed on Thursday

The FT, citing documents shared with OpenAI's backers, put the company's annualised revenue at approaching $50 billion as of the end of September. CNBC confirmed the figure through a person familiar with an investor presentation, which also touted 77% total run-rate growth in the third quarter and 107% run-rate growth for the enterprise business over the same period.

Those are strong numbers for any company on earth. They were still a shock, because two weeks earlier Axios had reported that OpenAI's annual recurring revenue was "nearing" $70 billion, and that figure had been repeated everywhere as settled fact, including by the people pricing Nvidia.

Where the $70 billion came from

The FT's explanation, from a person with knowledge of the matter: the discrepancy arose from attempts by OpenAI's own investors to produce a direct comparison with Anthropic's annualised revenues. The two companies compute the figure differently. Anthropic includes revenue from sales made through cloud partners such as AWS and Google Cloud, while OpenAI does not.

"Efforts to 'gross up' OpenAI's annualised revenue led to reports that the group's annualised revenue had hit $40bn in August. The company has since told investors its revenues have grown more than 70 per cent, leading to the $70bn figure." (Financial Times)

Do the arithmetic and the mystery evaporates. $40 billion times 1.7 is $68 billion, which is precisely the number CNBC says was "widely reported late last month." A grossed-up base, multiplied by a real growth rate, produced a headline that OpenAI never issued. CNBC's source said the same thing from the other side: the $68 billion included gross revenue from OpenAI's partners, which helps investors make a more direct comparison with Anthropic.

Full disclosure: this blog reported the $40 billion run rate in August. The FT now says that figure was itself a product of the gross-up exercise. Consider this the footnote we should have had then.

Gross versus net, in one dollar

Semafor laid out the mechanics back in April: when a customer buys $1 worth of tokens through a cloud partner, OpenAI counts its 20-cent cut as revenue, while Anthropic counts the whole dollar. The Information reported in March that Anthropic lists the providers' share as a sales and marketing cost, while OpenAI treats Microsoft as the primary provider for Azure sales. Both approaches comply with US GAAP. Which one a company may use turns on who controls the customer relationship and delivers the product, the principal-versus-agent test.

If that still feels abstract, picture two restaurants on the same delivery app. The diner pays $30. One restaurant books $30 of revenue and writes the app's $6 fee off as marketing. The other books the $24 it actually receives. Same meal, same cash in the till, different top line. Neither is lying; one just looks bigger on a leaked slide.

Semafor put the gap at up to $8 billion of annualized revenue in April, when both companies were in the $20-25 billion range. The partner percentages vary deal by deal, so the gross-up is not a fixed multiplier. That is exactly why a borrowed one broke at $50 billion of scale.

The market read the headline, not the method

CNBC's tally of Thursday's damage: Nvidia down 3%, Oracle nearly 6%, CoreWeave nearly 8%, AMD 4%, Broadcom 4%, Intel 5% and Super Micro nearly 5%. Quartz noted the Nasdaq 100 fell 1.4%. None of those companies reported anything. They fell because the largest single buyer of their output turned out to be about a quarter smaller than a number it never published.

The Hacker News thread on the FT story ran to 429 points and 300 comments, split between "this is a non-story, investors did bad arithmetic" and "this is the story: a trillion-dollar private company the whole market leans on, with no audited numbers in public." Both readings are correct, which is the uncomfortable part.

What OpenAI says comes next

Bloomberg, as relayed by Quartz, reports that OpenAI told investors it expects annualized revenue to reach or exceed $70 billion by the end of 2026, with enterprise as the primary engine. So the $70 billion did not disappear. It moved from a September fact to a December target, which is a different thing to lend against.

Those projections came up in fundraising talks. Bloomberg says OpenAI is negotiating $30 billion or more at a $1.4 trillion pre-money valuation, with Abu Dhabi's MGX among the funds in the conversation. CNBC describes the round as early-stage and driven by investor demand, with no term sheet finalized. OpenAI closed a $122 billion round in March at an $852 billion valuation, and CFO Sarah Friar told CNBC last week the company remains "very well capitalized." The IPO prospectus was filed confidentially in June with a 2027 debut in view; Sam Altman said in September that "right now would be an ill-advised moment to go public," in part because of safety concerns.

Anthropic, the comparison that started all this, told investors in August that its run rate hit $65 billion at the end of July and is reportedly seeking a $2 trillion valuation. Reuters, cited by CNBC, puts its 2025 revenue at $4.6 billion against a $42 billion net loss, from a leaked prospectus. Whichever lab you prefer, its biggest number is still a gross-versus-net footnote away from someone else's.

Why this matters if you build on these APIs

• Run rate is not a GAAP number. The Information says both companies annualize by taking four weeks of revenue and multiplying by 13. That is a snapshot of a good month, not audited revenue, and it moves with whatever the biggest customers did in those four weeks.
• Gross versus net is your problem too. If you sell through AWS Marketplace, an app store or a reseller, decide once whether you report gross or net, write the policy down and apply it to every number you ever publish. If you don't, someone else will gross you up for you, and they will not be around when the market corrects.
• Three questions for every leaked lab figure. Whose number is it, is it gross or net, and as of when. The $68 billion failed all three and still moved Nvidia.
• Pricing pressure is unchanged. OpenAI's 77% growth and 107% enterprise growth are real, as is Anthropic's $65 billion. Nothing about Thursday suggests either lab needs to raise API prices to hit its targets. The accounting changed, not the demand.
• The capex trade is keyed to private leaks. Nvidia, Oracle and CoreWeave repriced on a document none of their shareholders can read. Expect more Thursdays until at least one of these companies files a public 10-Q.

Key Takeaways

• OpenAI's annualized revenue was roughly $50 billion at the end of September, per the FT and CNBC, not the ~$70 billion reported two weeks earlier.
• The $68-70 billion figure was investors' gross-up of OpenAI's net revenue to match Anthropic's accounting, then grown by the 70%+ rate OpenAI reported; $40B times 1.7 is $68B.
• Semafor's April mechanics: $1 of tokens through a cloud partner is $1 of revenue at Anthropic and about 20 cents at OpenAI. Both methods are GAAP-compliant.
• Chip and cloud stocks fell 3% to 8% on Thursday (Nvidia, Oracle, CoreWeave, AMD, Broadcom, Intel, Super Micro) without reporting anything themselves.
• OpenAI now tells investors it expects $70 billion or more by the end of 2026 and is negotiating a $30 billion-plus raise at a $1.4 trillion pre-money valuation, per Bloomberg.
• If you sell through a marketplace, pick gross or net once and never mix them; the comparison trap that caught OpenAI's investors applies to a $20K MRR SaaS too.

Sources: CNBC, Financial Times, Semafor, Quartz (on Bloomberg), The Decoder, The Decoder (on The Information), TechTimes, Hacker News