OpenAI Told Investors 68 Billion Dollars. Nine Days Later It Said 50 Billion, and Three Suppliers Repriced.

On Thursday 8 October 2026, Nvidia fell 3 percent, Oracle fell nearly 6 percent, and CoreWeave fell nearly 8 percent, closing down 7.77 percent. No product shipped. No contract collapsed. No competitor launched. Three companies that sell servers and power to the AI industry repriced in a single session because one customer changed a number it had given investors nine days earlier.

The number that did it is worth stating precisely, because the dossier-grade version of this story gets it wrong. OpenAI had told investors its annualised revenue was approaching 70 billion dollars, and the figure that circulated widely was 68 billion. On 8 October it told them September annualised revenue was almost 50 billion.

That is not a loss of 18 billion dollars in sales. Nothing was lost. The measure changed.

According to reporting from CNBC and Axios, the higher figure included gross revenue from OpenAI's partners. OpenAI counts net revenue. The gross number exists so that investors can put OpenAI on the same footing as Anthropic, whose own tally includes cloud partner sales. The Financial Times reported the lower figure first; Reuters and CNBC confirmed it.

So the sequence is: to make itself comparable to a rival, OpenAI adopted a measure that flattered it by about 18 billion dollars. The market traded on that measure for nine days. Then the comparison basis was dropped, and three suppliers were repriced.

That is not fraud. It is something more ordinary and more instructive, which is that nobody had agreed on what the word meant.

Three revenues, one company#

Here is the part that should stick with you. There are at least three different quantities being called revenue in circulation for OpenAI, and the spread between them is larger than most of the growth being reported.

The first is actual annual revenue. OpenAI reported 13.07 billion dollars of revenue for 2025, alongside a net loss of 38.5 billion.

The second is annualised run rate. This is what the CFO publishes as the headline growth series. In her January 2026 post, Sarah Friar wrote that revenue grew roughly three times year over year, "or 10X from 2023 to 2025: $2B ARR in 2023, $6B in 2024, and $20B+ in 2025." An annualised run rate takes a recent month's revenue and multiplies it by twelve. It is a legitimate measure and it is not the same thing as money received in a calendar year.

The third is gross partner-inclusive revenue. That is the 68 billion that got the market moving.

So for 2025: 13.07 billion actual, 20 billion-plus annualised. A gap of more than half, for the same company, in the same year, describing the same thing. Neither number is false. They answer different questions, and almost nobody states which one they are quoting.

Then in September 2026 a fourth figure appeared, and it turned out to be a fifth definition. The number was constructed for comparability with a competitor.

The compute claim has the same shape#

Now apply the same discipline to the physical capacity claim, because it has an identical structure.

In January 2025, with the Stargate announcement, OpenAI committed to securing 10GW of AI infrastructure in the United States by 2029. In an April 2026 company post titled Building the compute infrastructure for the Intelligence Age, OpenAI said it had met that target more than a year early and had added 3GW in a single quarter.

Four months before that announcement, on 18 January 2026, the CFO had published this:

"Compute grew 3X year over year or 9.5X from 2023 to 2025: 0.2 GW in 2023, 0.6 GW in 2024, and ~1.9 GW in 2025."

Read those two documents side by side. In January, live compute was about 1.9GW. In April, the company said it had surpassed 10GW secured. Both are OpenAI. The word doing the work is secured, and securing capacity and running capacity are genuinely different things. A contract with a cloud provider is not a live cluster.

The reporting on the April post picked this up. Data Center Dynamics, covering the same announcement, noted it was likely contracted rather than live capacity, and cited the CFO's 1.9GW end-2025 figure as the last disclosed live number.

OpenAI's own blog post does not draw the distinction. A reader comparing the company's two documents can do it without any external claim, which is the whole point.

The target has since moved again. OpenAI said in April 2026 that it was now planning for 30GW of compute by 2030. Thirty gigawatts is roughly sixteen times the compute the entire company had live at the end of 2025.

The commitment figure will not sit still#

The spending story has the same instability, and this part is genuinely a fight inside the company rather than a definitional dispute.

In February 2026, OpenAI told investors it was now targeting roughly 600 billion dollars of total compute spend by 2030. That came months after Sam Altman had touted 1.4 trillion dollars in infrastructure commitments. By July, reported planned cloud spending was at 750 billion dollars.

Three numbers, and they are not a clean escalation. The 1.4 trillion was an eight-year figure. The 600 billion and the 750 billion are four-year figures. Setting them beside each other without normalising the horizon would be a category error, and it is an error that gets made constantly.

OpenAI has publicly denied there is a rift between the CEO and the CFO over compute spending. The Wall Street Journal has reported otherwise. The company's own statement is that there is "no truth" to the idea of a rift.

What is not in dispute is that the commitment is large relative to the revenue, and that OpenAI is buying most of it as cloud commitments rather than owning sites. That is what keeps the balance sheet light, which is precisely what the CFO says the strategy requires.

The counterweight, in the CFO's own words#

Here is the part that cuts against the standard "Silicon Valley recklessness" framing, and it deserves stating at full strength.

Friar's own published description of the financial model is one of deliberate structure. On the compute portfolio, she wrote:

"Compute has scaled at roughly 3x per year"

and on managing the gap between capacity and usage:

"We manage that by keeping the balance sheet light, partnering rather than owning, and structuring contracts with flexibility across providers and hardware types. That lets us lean forward when growth is there without locking in more of the future than the market has earned."

Her stated focus for 2026 is "practical adoption". She also noted the diversification directly: "Three years ago, we relied on a single compute provider."

None of that is reckless. It is a coherent argument that you should buy capacity in stages, keep commitments cancellable, and diversify across providers and hardware. If that argument is true, the risk in this story is concentration and timing, not imprudence.

There is a real tension, though, and it is not resolved anywhere in the public record. A model built on buying in tranches against demand, and a roughly 750 billion dollar multi-year commitment, sit together comfortably only if the commitment is heavily back-ended and genuinely cancellable. The contracts are with AWS, Oracle, Microsoft and CoreWeave, and their counterparty credit terms and cancellation clauses are not public. Nothing a reader can see establishes which of those two things is true.

Why the market reacted to an accounting note#

The supplier moves are the part that should worry people who run infrastructure rather than funds.

CoreWeave sells GPU capacity to OpenAI. Oracle sells Oracle Cloud Infrastructure to OpenAI. Nvidia sells the chips under all of it. When a customer of that size changes what it reports, the customers reprice, and then the suppliers reprice. CoreWeave fell nearly 8 percent in a day on this, closing at 81.58 dollars.

Look at the mechanism carefully. A number that was constructed partly for competitive comparability travelled outward to investors, moved three listed companies, and was then partly retracted nine days later. Nobody had to be wrong about anything for real money to change hands. The only thing that happened is that a definition was disclosed.

That is the transferable lesson, and it has nothing to do with OpenAI specifically. Any company that reports a run rate, a gross number, or a partner-inclusive figure is making a choice about definition. The choice is rarely dishonest. It is almost always unannounced.

Where this could be wrong#

The obvious objection is that this makes a mountain out of an accounting clarification that was handled correctly and quickly.

The defence is fair on the narrow point. Nine days is a fast correction. CNBC confirmed the lower figure independently rather than taking the company's word for it. The stated reason was specific and verifiable. None of that is concealment.

But three limits deserve stating.

First, the live compute figure is nine months stale. The 1.9GW number is from January 2026. OpenAI may have substantial live capacity now that it has not published. The gap is between published figures, not proven idleness. Anyone who tells you the 10GW claim is false is overstating it as confidently as anyone who tells you it is fine.

Second, OpenAI is private. There are no SEC filings, no audited financial statements, and no third-party audit of any capacity or revenue figure in this post. Every number here is either a company statement or press reporting of a company statement. I have not independently verified a single financial figure in this piece, and the distinction between a primary document and a sourced report is doing real work throughout.

Third, the 2025 loss figure is the weakest number here. The 38.5 billion net loss on 13.07 billion of revenue is press reporting and should be treated as indicative rather than confirmed. The 2030 revenue target of roughly 280 billion is also press reporting of what the company has told investors, and a target stated to investors by a pre-IPO company is not a forecast in any rigorous sense.

Finally, the comparison with Anthropic is doing a lot of work in the public telling and I cannot fully evaluate it. Anthropic's annualised revenue is reported around 65 billion dollars. Whether the two companies' revenue is now measured on compatible bases is precisely the question the 8 October correction reopened.

The closing note#

OpenAI's disclosures contain enough to check each claim against another of its disclosures. That is worth stating because it is not true of most companies at this scale, and it is the reason this post exists without a single leaked document.

The company's own CFO published a compute series ending at 1.9GW. Four months later the company announced it had passed 10GW. The company's own investors were told 68 billion, then told 50 billion. Both sequences are checkable from public documents, by a reader with no special access.

Whether that is a company being loose with language, or a business running on several incompatible definitions at once, is the question. The public record does not settle which, and no amount of additional press reporting will. The documents do, and the documents are all public.

The part worth carrying into next week: when a number moves a market, find out what it measured before you decide what it means.