A number from the witness box

The most useful disclosure to come out of Musk v. Altman this week was not about the founding emails, or the side letters, or whether Sam Altman is, in some doctrinally meaningful sense, the keeper of the original mission. It was a number Greg Brockman said in open court, under oath, in Oakland.

OpenAI expects to spend $50 billion on compute in 2026.¹

Brockman framed it as a growth statistic. In 2017 the company spent $30 million on compute; in 2026 it will spend roughly 1,667 times that. He offered the comparison, I think, to make a point about how far the work has come. It also makes a different point, which is that a private company with no profits is about to spend more on a single line item than the entire annual revenue of Lockheed Martin.

What was actually said. Brockman was on the stand as part of Musk's ongoing claim that the conversion of OpenAI's commercial arm undermined the founding charitable purpose. The $50bn figure came up in the context of explaining why the company needs the capital structure it has, which is a reasonable thing to put in front of a judge if your defence is that the non-profit model was never going to fund what the work requires. It is also the most concrete forward-looking compute number any frontier lab has put on a record under oath. Microsoft's Azure disclosures imply a number in this range for OpenAI's allocation; analyst estimates have circled it; the Stargate announcements gestured at it. But Brockman said it, on a witness stand, as a planning figure.

The frame this fits. Inference economics is the frame that does most of the work here, but it shares the load with AI performativity. On inference: the binding cost has shifted from training the next model to serving the current one, and $50bn is not a training budget. Even a Stargate-scale training run, run repeatedly across a year, doesn't get you to fifty. The number is dominated by serving load, ChatGPT, the API, the enterprise deployments, the agentic products that turn a single user request into many model calls. This is what the frame predicted: as models get embedded into workflows, inference cost stops being a back-of-envelope rounding error and becomes the company.

On performativity: $50bn of committed compute is itself a market signal, independent of whether the underlying revenue justifies it. It commits Nvidia's order book, it commits hyperscaler capex, it commits the power-purchase agreements behind the data centres, and it commits OpenAI to a revenue trajectory steep enough to make the spend look retrospectively sensible. Capital at this scale shapes the market it is allegedly responding to.

Where the numbers strain. OpenAI's most recently reported annualised revenue run-rate is in the $13bn range as of mid-2025, with internal projections that have been variously leaked at $11.6bn, $12.7bn, and figures north of $20bn by year-end 2026.² Even taking the most generous projection, call it $30bn ARR exiting 2026, the company is planning to spend $50bn on compute against it. That is not a gross margin problem; that is a structural one. The standard defence is that compute spend is partly capex-like, amortised against future revenue, and that a chunk of the $50bn is effectively a pre-purchase of capacity that will serve 2027 and 2028 demand. This is plausible. It is also exactly the argument that every infrastructure-heavy company has made before the cycle turned.

The standard defence is that compute spend is partly capex-like. This is plausible. It is also the argument every infrastructure-heavy company has made before the cycle turned.

What this is a case of. This is the disclosure that previously came out as estimates from semis analysts, leaks to The Information, or implied numbers tucked inside Microsoft's MD&A. The shift is that it is now on a court record, which means it is the working planning figure inside OpenAI as of Q1 2026, not a stretch case, not a marketing number, not something a founder said at a conference. The trial venue is doing what trial venues do, which is converting commercial opacity into evidentiary clarity. (The same testimony, incidentally, produced an admission from Brockman that he never paid out a pledged $100,000 personal donation, a smaller and funnier disclosure, but a reminder that what gets said under oath is not always what the witness expected to be saying.)

What to watch. Three things. First, Microsoft's next 10-Q, which will show whether the Azure capex line is being built consistent with $50bn of OpenAI consumption, Microsoft cannot deliver this compute without spending against it, and the timing has to show up. Second, whether OpenAI's next disclosed revenue figure, formal or leaked, materially closes the gap to the spend; if ARR is still in the high teens by mid-2026, the financing structure has to do work that revenue isn't doing. Third, whether other labs respond with their own forward compute disclosures. Anthropic, in particular, has been quieter about its compute plans than the underlying Amazon and Google commitments suggest it can afford to be. If $50bn becomes the new floor of credibility for being a frontier lab, the labs that can't say a number that large have a different problem than the ones that can.

The frame fits. The number is the number. The question is what revenue does next.

Footnotes

¶