OpenAI's missing $20 billion belongs in your next AI contract

For about nine days this autumn, the number everyone in AI procurement quoted was $70 billion. On 29 September Axios reported, citing people familiar with OpenAI's financials, that its run rate had climbed to nearly that and that enterprise sales had more than doubled since July. On 8 October the Financial Times reported, citing documents shared with investors, that OpenAI's annualised revenue was about $20 billion lower than had been signalled. CNBC confirmed OpenAI had told investors it was at roughly $50 billion at the end of September. Oracle fell nearly 6% that day and CoreWeave nearly 8%.

I expected this to be a story about chip stocks and IPO valuations. What I keep thinking about is the customer. According to Sacra, enterprise is more than half of OpenAI's revenue. Those are the companies whose CIOs standardised on a model, retrained staff around it and signed renewals partly because the vendor looked like it was compounding faster than anyone. The growth figure was a sales argument as much as an investor one, and enterprise buyers were on the receiving end of both.

OpenAI's defenders have a decent case, and I want to give it its due. OpenAI never published $70 billion. The FT says its own investors built that figure by taking an August estimate of around $40 billion and adding the 70% growth OpenAI had reported, in an attempt to line it up against Anthropic. CNBC's source says the $68 billion version included gross revenue from partners, and TechCrunch points out that Anthropic counts sales through its cloud partners while OpenAI does not. Even $50 billion is up from roughly $20 billion at the end of 2025. OpenAI raised $122 billion in March, and CFO Sarah Friar said last week the company remains "very well capitalized".

I accept every word of that, and I think it strengthens the buyer's case. If a number assembled by OpenAI's own backers could circulate as fact for over a week at the most scrutinised private company on earth, the run-rate slide your account executive shows you deserves no more trust. Each vendor counts revenue its own way and picks the definition that flatters it in front of IPO investors. A company that closed $122 billion in March and is already in early talks about another $30 billion is setting prices with its next round in mind.

Anthropic deserves the same scepticism. It told investors its run rate hit $65 billion at the end of July and is reportedly seeking a $2 trillion valuation. Reuters, citing a leaked copy of its prospectus, says 2025 revenue was $4.6 billion against a net loss of $42 billion. New Constructs, an independent research firm, puts the company's value at $150 billion. Whichever number you believe, a vendor losing that kind of money has every incentive to sign logos cheaply now and recover margin once switching costs bite.

Roadmaps carry the same risk. OpenAI recently pulled its planned GPT-6.1 Astra launch, saying the model did not meet its safety standards. That may well have been the right call. Any buyer whose 2027 plan assumed Astra would ship now has a gap in that plan, and I have yet to see an enterprise AI contract that pays the customer anything when a promised model is withdrawn.

The ugly end of this is Builder.ai. After the AI app builder collapsed, it cut its FY24 adjusted revenue from $220 million to $55 million and its FY23 figure from $180 million to $45 million. A creditor restricted access to nearly $40 million in debt funding, the company was left with $5 million and it laid off all its staff. I do not think OpenAI or Anthropic is heading there. Builder.ai does show that a private AI vendor's revenue can be restated by three quarters after the fact, and the customers whose products ran on it found out at the same time as everyone else.

So stop negotiating AI contracts as though the vendor's pitch deck were an audited filing. Every agreement should give you the right to terminate with a pro-rata refund of prepaid fees if the vendor retires the model you contracted for, restates its revenue or raises at a lower valuation than its last round. It should also guarantee export of your data, prompts and fine-tuned weights in a format a rival can actually load. This week, ask procurement to pull every AI contract signed in 2026 and mark which ones let you walk away if the model you bought disappears; any that do not should be the first renewal you reopen.

Prompted by Nvidia, Oracle, CoreWeave and other AI stocks sink on OpenAI revenue report, CNBC.