OpenAI's compute line climbed to $856bn on other people's balance sheets

Someone will sit in a procurement review next week with a multi-year OpenAI commitment in front of them and these July figures in the appendix. The burn is the line that gets read aloud in that room; the revenue assumption underneath it decides whether the contract price holds.

The 2030 revenue figure is a factor of about 9.7 on this year's ^[4]. Free cash flow is the residual of that assumption, spending minus revenue, and the spending is largely contracted while the revenue is the part that has to arrive, so a miss on revenue does not shave the burn, it adds to it ^[12].

The two lines moved in opposite directions because much of the build is financed by partners instead of by OpenAI ^[18]. Oracle is spending more on data centres than it earns in a quarter, much of it against OpenAI commitments, and that capital expenditure lands on Oracle's accounts ^[8]. SB Energy took $5.5bn in OpenAI warrants to sign a 20-year lease ^[9]. That converts a capital commitment into an operating one and pays for it in equity instead of cash ^[9].

What the financing costs is visible in one deal. Oracle needed PIMCO to anchor $10bn of a $16.3bn data centre financing after US banks stepped back ^[10], roughly 61% of the total from a single asset manager ^[5].

Some of the roughly $256bn gap ^[1] may be definitional: February's number was described as compute, July's as computing power and infrastructure ^[13]. And these are projections in a document prepared to win a computing deal, not audited accounts, already revised between May and July ^[14]. Bloomberg, which published the figures in a write-up by Samantha Oltman, corrected its story shortly after publication ^[15].

What governs a customer's own contract is narrow: which legal entity owns the capacity they are buying, and who is owed money against it. Free cash flow answers neither question, because it measures money leaving one entity and not obligations created ^[11]. I think the entity map is the thing to ask for before signing a multi-year commit. Expect a refusal: counterparty terms in structures like the SB Energy warrants are not a customer's to inspect. A shorter term at a worse headline rate is the realistic fallback. It prices better against a vendor that the FT reports is on track to exhaust its March raise by 2028, while in talks valuing it at about $1.2 trillion ^[16][17].