OpenAI Disclosed a $50 Billion Revenue Run Rate. AI Stocks Fell Because Investors Had Priced In $68 Billion.

$50B (OpenAI's own disclosed revenue run rate, end of September -- not $68 billion) OpenAI told investors its annualized revenue run rate was approximately $50 billion at the end of September, the Financial Times reported Oct. 8, with CNBC independently confirming the figure. That's about $18 billion below the $68 billion run rate that had circulated among investors for weeks -- a number built entirely from anonymously sourced estimates and never an audited figure. Oracle, Nvidia and CoreWeave led a broader AI-infrastructure stock selloff on the news, with Oracle down as much as 6% intraday and CoreWeave off more than 7%; the Nasdaq fell roughly 1.4% for the session amid broader market weakness.

The $68 billion figure traces back to a chain of leaked estimates, not a single report. Bloomberg reported a roughly $40 billion run rate in mid-August, citing people familiar with OpenAI's financials. OpenAI then told investors its revenue had grown roughly 70% in the weeks after, and Axios reported Sept. 29 that the run rate was "nearing $70 billion." The arithmetic is internally consistent -- $40 billion grown 70% is $68 billion -- even though neither the $40 billion base nor the 70% growth claim was ever an official OpenAI number. (An "annualized revenue run rate" isn't a GAAP figure. It's a snapshot of recent revenue multiplied out across a year, which is exactly why two companies measuring it slightly differently can both be technically accurate and still not comparable.)

What each number is actually counting

The explanation several outlets settled on, each citing a person familiar with the disclosure, is that the $68 billion figure was never OpenAI's own number -- it was investors' attempt to make OpenAI's revenue comparable to Anthropic's, by adding in gross revenue from OpenAI's reseller and platform partners the way Anthropic's own $65 billion figure is said to already do. OpenAI's $50 billion disclosure, by contrast, appears to count only revenue OpenAI itself collects directly. If that's right, the two companies' real run rates may be closer than a flat $50-billion-versus-$65-billion comparison suggests -- because nobody has published what Anthropic's figure looks like on OpenAI's narrower basis, or what OpenAI's figure looks like on Anthropic's broader one. Stacking two numbers that were never measured the same way and calling the gap a finding is exactly the kind of comparison this episode should make readers distrust.

A measurement correction moved markets anyway

Thursday's stock reaction is a little strange once you sit with it: OpenAI's own disclosed growth rates -- 77% total, 107% enterprise -- aren't bad numbers, and nothing in Thursday's reporting says OpenAI's business actually slowed down. What moved is the market's mental model of how big that business already is. Oracle, Nvidia, Microsoft and CoreWeave have all signed multi-year compute contracts with OpenAI worth tens of billions of dollars, priced against assumptions about how large and fast-growing OpenAI's revenue would be. A $50 billion base compounding at 77% a year reaches a very different place, several years out, than a $68 billion base compounding at the same rate. Thursday's selloff reads less like a reaction to today's number than a repricing of every multi-year spending commitment built on top of yesterday's wrong one.

Who actually has something at stake here

OpenAI is widely expected to go public in early 2027, having confidentially filed a draft prospectus with the SEC in June, at a valuation investors have put as high as $852 billion. Anthropic filed its own confidential S-1 the same month and has been the subject of IPO-timing reports all autumn. For now, both companies are still describing their own size to the public exclusively through investor leaks and conference-call color -- not through the audited, reconciled numbers a public prospectus actually requires. Nothing about OpenAI's underlying business changed between Wednesday and Thursday. What changed is that a number investors had been treating as confirmed turned out to be reverse-engineered from two different anonymous leaks, run through arithmetic nobody at OpenAI had actually checked.