The $20 Billion That Was Never There
OpenAI did not lose revenue. The comparison did the work — and that is the more useful story for anyone buying AI.
The interesting thing about OpenAI’s missing $20 billion is not that the money vanished. It is that it was never OpenAI’s number in the first place.
According to the Financial Times, reporting on financial documents shared with investors, OpenAI has told its backers that its annualised revenue was approaching $50 billion at the end of September. A little over a week earlier, the figure circulating in reports — the FT’s own among them — was approaching $70 billion. Same company, same month, a $20 billion difference, and no collapse in demand required to explain it.
That sounds like a revenue miss. It isn’t. It is a measurement problem that got priced as a growth problem.
Two run rates, two definitions
Annualised revenue is not revenue. It is a snapshot of a recent period stretched out to a year, and it only means anything if you know what was counted in the snapshot.
The FT reports that the higher figure came from attempts by OpenAI’s own investors to produce a direct comparison with Anthropic’s annualised revenue. The two companies do not count the same thing. Per the FT’s reporting, Anthropic includes revenue from sales made through its cloud partners, such as Amazon’s AWS and Google Cloud. OpenAI does not include partner sales that way.
So the $70 billion was a grossed-up translation. It took OpenAI’s number and restated it in Anthropic’s accounting language so the two could sit in the same chart. Somewhere between the investor deck and the headline, the translation became the original.
The underlying growth, on the figures the FT cites from the investor documents, is still steep. The same reporting puts OpenAI at close to $30 billion in annualised revenue in July, rising to approaching $50 billion by the end of September — growth of more than 70% in a single quarter. TechCrunch, summarising the FT report, also notes the context that makes the run rate matter so much: OpenAI raised $122 billion in a March funding round alone, per TechCrunch, and its leaked 2025 financials, reported earlier this year, showed about $13 billion in revenue against significantly higher spending. A company spending at that scale does not get to have a quiet rounding difference.
Why the market treated it as news anyway
The market reaction tells you how much weight this one metric is carrying. Following the FT’s report on Thursday, US tech stocks fell, with the Nasdaq 100 closing down 1.4%, Nvidia down 2.9%, Oracle down 5.5%, and Micron down 4.8% — figures reported by the FT. Nothing about chips, cloud contracts, or end demand changed that afternoon. A definition changed, and billions in market value moved with it.
That is the part worth sitting with. The annualised revenue of two private companies has become a proxy for the demand for an entire industry’s infrastructure build-out. When the proxy is built from two incompatible definitions, the error does not stay inside a spreadsheet. It propagates into public share prices.
There is also a competitive reading that the correction makes sharper, not softer. CNBC reported that Anthropic’s annualised run rate reached $65 billion by late July, a figure cited by Mobile World Live in its coverage of the FT story. If Anthropic counts partner sales and OpenAI does not, a headline comparison between the two numbers was never a comparison. Anyone ranking the labs by run rate this autumn has been ranking accounting policies at least as much as businesses.
The catch for builders
None of this requires OpenAI to be in trouble, and the honest version of the story should not pretend otherwise. Approaching $50 billion annualised, up more than 70% since July on the FT’s figures, is not a weak business. The IPO timing adds a second caveat: TechCrunch reports that an OpenAI listing once rumoured for this year has been pushed to early 2027. A private company can survive a confused metric. It has no quarterly earnings call where it has to spend the first twenty minutes explaining what a run rate is.
The practical problem lands on the buyers. If you are purchasing AI capacity — API seats, enterprise contracts, reserved inference — you are making decisions inside a market whose headline demand signal just moved $20 billion without a single customer cancelling. Vendor comparisons, pricing negotiations, and capacity promises all lean on run-rate claims that turn out to be definition-dependent. The model providers know which definition flatters them. The invoice, notably, does not use a run rate. It uses your tokens.
So the useful discipline here is boring: when a lab quotes a revenue figure, ask what is counted in it. Partner sales or direct only. Gross or net of revenue share. A month annualised, or a quarter. The answers change the number by tens of billions at the top of this market, which means they change it by proportionally similar amounts in the smaller claims you hear in a sales meeting.
For builders, the takeaway is practical: treat every run rate as a claim about a definition before you treat it as a claim about demand. OpenAI’s September did not get $20 billion worse last week. The chart did.
That is the problem.
Source: OpenAI’s revenue is reportedly $20 billion less than previously projected