Roughly $50 billion. That’s the annualized revenue figure attached to OpenAI in a CNBC report on Thursday, and it’s about $18 billion shy of the $68 billion gross number that had been circulating widely. The difference wasn’t a collapse in demand or a canceled contract. It came down to how partner revenue gets counted.
Nvidia, Oracle, CoreWeave, and other AI names sank on the news. Neither report specified how far they fell or on what volume, so I’m not going to pretend I know. What I do know is that an accounting distinction moved a whole category of stocks in a single session, and that’s the part worth sitting with.
Gross versus net is not a technicality
I review AI tools for a living. The single most common thing I run into isn’t bad software. It’s numbers that are technically true and practically useless. A vendor tells me their agent framework handles “10,000 requests per second” and it turns out that’s cached reads on a warm instance with no tool calls. Another tells me their coding assistant has a 90% acceptance rate, which means 90% of suggestions weren’t immediately dismissed, not that 90% shipped.
Gross versus net revenue is the same genre of problem, just with more zeros. When revenue flows through partners, you can present the full amount passing through the system or the portion the company actually keeps. Both are real figures. They describe very different businesses. If the market had been pricing AI infrastructure off the larger number and the smaller one is closer to what lands, that’s a repricing event, not news about OpenAI’s health.
The thing that should bother you isn’t the gap. It’s that the gap persisted long enough to become consensus.
Why this hits infrastructure stocks specifically
Nvidia sells the chips. CoreWeave rents the compute. Oracle provides capacity. None of them are OpenAI, but all of them are priced partly on assumptions about what companies like OpenAI can afford to spend over the next several years. Spending capacity is downstream of revenue. Shrink the revenue assumption and the spending assumption follows.
That’s the mechanism. It’s also why the reaction was broad rather than targeted. Traders weren’t making a judgment about any one company’s quarter. They were adjusting an input that feeds into a lot of models at once.
What I take from this as a reviewer
There’s a lesson here that applies well below the level of public markets, down to the daily work of picking tools for your team.
- Ask what’s being counted before you ask how much. A number without a definition is a vibe. Revenue, latency, accuracy, token cost, uptime — every one of these has a generous version and an honest version.
- Watch for revenue that passes through. If a tool’s pricing is built on reselling someone else’s model access, you’re exposed to two margin structures, not one. That matters when either side adjusts.
- Treat widely cited figures with more suspicion, not less. The $68 billion number got repeated because it got repeated. Consensus is not verification.
- Separate the company from the category. OpenAI disclosing a smaller net figure says something about accounting presentation. Whether it says something about AI demand is a separate question that this report doesn’t answer.
What we actually learned
Not much about AI adoption, honestly. The reports didn’t describe falling usage, churn, or missed product targets. They described a clearer picture of how one company’s revenue is structured. Everything else was the market doing arithmetic in public.
I find that genuinely useful, though, in a way that’s easy to miss. For two years the AI buildout has been financed on projections, and projections are only as good as the figures feeding them. A correction that comes from better disclosure is the healthy kind. It’s the alternative — assumptions compounding quietly until something breaks — that should worry people.
If you’re evaluating AI tools right now, the practical takeaway is small and boring and I’d stand behind it anyway. Read the methodology. Ask the vendor which number they’re giving you and what it excludes. When a figure sounds too clean, it usually is, and the cost of asking is one email while the cost of not asking shows up later, at scale, in your invoice.
Eighteen billion dollars of difference between two defensible presentations of the same business. That’s not a scandal. It’s a reminder that the number on the slide is a choice someone made, and you’re allowed to ask why they made it.
🕒 Published: