The Take — OpenAI's $20B gap is definitional, and that's worse

Share
The Take — OpenAI's $20B gap is definitional, and that's worse

The $20 billion never went missing from OpenAI's business — it was never in it. OpenAI's annualized revenue was always a number only OpenAI gets to define, and with a confidential 2027 IPO filing on record and a $1.2 trillion private round under consideration, I think a self-defined metric heading into underwriter season is worse than a number that was simply wrong. A wrong number gets corrected once; a self-defined number survives every headline it produces.

Our afternoon brief on Wednesday laid out how the correction landed. The Financial Times broke the story Thursday: OpenAI told investors its annualized revenue was approaching $50 billion at the end of September, against the roughly $70 billion that had circulated in investor documents and media coverage. Part of the gap is definitional rather than demand-driven — CNBC reported that the earlier $68 billion figure included gross revenue from OpenAI's partners, money OpenAI itself never books, and the $70 billion framing took hold anyway. The market's first answer was a shrug priced in red: Nvidia gave up about 3 percent, Oracle dropped 6 percent, CoreWeave slid 8 percent, with AMD, Broadcom, Intel and Super Micro also in the red.

Here is why I think the definitional framing is the alarming part, not the comforting one. This is the second time in roughly two months that a run-rate figure has been the story rather than the business. In August we reported OpenAI's revenue run rate tops $40B ahead of IPO — a number that came, like this one, from the company's own deck. Run-rate revenue is not audited, not standardized, and not even consistently defined across the labs that quote it: it extrapolates one month, includes whatever the issuer decides counts as revenue, and stops there. The whole industry now prices itself on it — Arena, which just raised $200 million at a $3.1 billion valuation, told investors its annualized revenue passed $100 million in June — but only Arena, and only OpenAI, can tell you what sits inside that number. Until a filing forces the question, nobody else can check it.

The counter-case is real and I take it seriously. Nothing here looks like a demand problem. The same investor materials showed 77 percent run-rate growth in the third quarter and 107 percent growth on the enterprise side, according to a person familiar with the deck who spoke to CNBC. Gross-versus-net treatment is a normal, defensible accounting choice — platform businesses report partner volume all the time — and the $70 billion figure that panicked the tape was largely a media embellishment of a $68 billion number that itself was never OpenAI's books. On this reading, Thursday was a correction of a rumor, not of a business: growth of 77 percent is not what a collapsing company posts, and if the lab were quietly missing its numbers, the enterprise line would not be doubling. The bulls will also note the market has months to learn the difference before any roadshow.

Why the take holds anyway: because the correction is the proof. If the metric were routine, correcting it would have moved nothing. Instead one clarifying sentence took roughly $20 billion of headline revenue off the table and pulled Nvidia, Oracle and CoreWeave down with it — which tells you exactly how much valuation weight the number was already carrying. And the ambiguity does not vanish now that it has been explained; it compounds. The IPO was filed confidentially with a 2027 target, which means every run-rate figure OpenAI publishes between now and then is an unaudited, issuer-defined claim that the market will price anyway. We dug into the underlying economics last month — OpenAI's cash gap shrank while its compute bill grew — and the pattern is the same shape: the lab's story keeps improving on metrics the lab computes, while the obligations that will be scrutinized in a prospectus keep growing on metrics auditors compute. The leak-and-refute cycle around these figures is quietly becoming its own risk factor: a company that lets a flattering number circulate and corrects it after markets move is not lying, but it is learning that the correction costs less than the clarification would have.

What would change my mind: standardization before the roadshow. If OpenAI — or any frontier lab — publishes restated historical revenue under a defined standard, with third-party assurance, before it talks valuation again, the definitional objection dies on the spot: then run-rate is a comparable metric instead of a mood. So would evidence that underwriters are already forcing GAAP-style disclosure inside the confidential filing; if the audited numbers are in there and the run-rate chatter is just marketing, I am worried about the wrong thing. Until either shows up, I would treat every lab's run-rate the way the market briefly treated $70 billion on Thursday: as a claim, not a fact — and note that this time, nobody had to fabricate anything for it to be wrong.

If the lab defines the metric and the market prices on it, who exactly checks the number? Tell us in the comments.

Read more

OpenAI's first Category 5 influence op targeted editors, not feeds

OpenAI's first Category 5 influence op targeted editors, not feeds

OpenAI banned two state-linked influence campaigns on October 8 — and the number worth sitting with is not the ban count but the rating attached to one of them: the first Category 5 operation the company has disrupted in two and a half years of publishing threat reports. The deeper signal, though, is in the fine print of what the models were actually used for. What happened OpenAI's report describes two operations it calls "false front" entities — shells that launder geopolitical messaging

OpenAI busts influence ops that planted fake stories in real media

OpenAI busts influence ops that planted fake stories in real media

The day's AI news runs through one seam: the work is showing up in places nobody planned for — inside real newsrooms, across the whole night sky, and in the M&A column. OpenAI has banned two state-backed influence operations that used ChatGPT to plant fabricated stories inside legitimate news outlets — and rated the Russian one the most disruptive it has seen in two and a half years. In a report dated October 8, OpenAI detailed "Dark Clark," run from Russia across Latin America, which ran a th

Open Source Radar — October 9: plugins, sandboxes, tokens

Open Source Radar — October 9: plugins, sandboxes, tokens

Today's open-source signal is infrastructure rather than hype: Microsoft's code sandbox reaches 1.0, Anthropic's knowledge-worker plugins keep climbing, a beloved token counter flips its default, and LocalLLaMA squeezes a usable 2B model into about 700 MB. knowledge-work-plugins (Python, ~27,900 stars, Apache-2.0) — Anthropic's repository of role-shaped plugins for Claude Cowork is the top AI repository on today's daily trending page, and the stars keep coming: roughly 2,100 more than when we

Deep Dive — The four-token blind spot inside DeepSeek V4

Deep Dive — The four-token blind spot inside DeepSeek V4

ByteDance's Seed research team says it has found the cause of one of the stranger recurring complaints about DeepSeek's models: the same question, asked with nothing changed except a few junk characters bolted onto the front, can flip the model from right to wrong. Their paper, posted to arXiv on September 28, traces the wobble to a memory-saving trick used during long-context inference, and reports that DeepSeek-V4-Flash-Base's retrieval accuracy swings by as much as 40.2 percentage points depe