Soundtrack: Queens of the Stone Age - Song For The Dead
Editor's Note: The original piece had a mathematical error around burnrate, it's been fixed.
Also, welcome to another premium issue! Please do subscribe, this is a massive, 7000-or-so word piece, and that's the kind of depth you get every single week for your subscription.
A few days ago, Sam Altman said that OpenAI’s revenues were “well more” than $13bn in 2025, a statement I question based on the fact, based on other outlets’ reporting, OpenAI only made $4.3bn through the first half of 2025, and likely around a billion a month, which I estimate means the company made around $8bn by the end of September.
This is an estimate. If I receive information to the contrary, I’ll report it.
Nevertheless, OpenAI is also burning a lot of money. In recent public disclosures (as reported by The Register), Microsoft noted that it had funding commitments to OpenAI of $13bn, of which $11.6bn had been funded by September 30 2025.
These disclosures also revealed that OpenAI lost $12bn in the last quarter — Microsoft’s Fiscal Year Q1 2026, representing July through September 2025. To be clear, this is actual, real accounting, rather than the figures leaked to reporters. It’s not that leaks are necessarily a problem — it’s just that anything appearing on any kind of SEC filing generally has to pass a very, very high bar.
There is absolutely nothing about these numbers that suggests that OpenAI is “profitable on inference” as Sam Altman told a group of reporters at a dinner in the middle of August.
Let me get specific.
The Information reported that through the first half of 2025, OpenAI spent $6.7bn on research and development, “which likely include[s] servers to develop new artificial intelligence.” The common refrain here is that OpenAI “is spending so much on training that it’s eating the rest of its margins,” but if that were the case here, it would mean that OpenAI spent the equivalent of six months’ training in the space of three.
I think the more likely answer is that OpenAI is spending massive amounts of money on staff, sales and marketing ($2bn alone in the first half of the year), real estate, lobbying, data, and, of course, inference.
According to The Information, OpenAI had $9.6bn in cash at the end of June 2025.
Assuming that OpenAI lost $12bn at the end of calendar year Q3 2025, and made — I’m being generous — around $3.3bn (or $1.1bn a month) within that quarter, this would suggest OpenAI’s operations cost them over $15bn in the space of three months. Where, exactly, is this money going? And how do the numbers published actually make sense when you reconcile them with Microsoft’s disclosures?
In the space of three months, OpenAI’s costs — if we are to believe what was leaked to The Information (and, to be clear, I respect their reporting) — went from a net loss of $13.5bn in six months to, I assume, a net loss of $12bn in three months.
Though there are likely losses related to stock-based compensation, this only represented a cost of $2.5bn in the first half of 2025. The Information also reported that OpenAI “spent more than $2.5 billion on its cost of revenue,” suggesting inference costs of…around that?
I don’t know. I really don’t know. But something isn't right, and today I'm going to dig into it.
In this newsletter I'm going to reveal how OpenAI's reported revenues and costs don't line up - and that there's $4.1 billion of cash burn that has yet to be reported elsewhere.