OpenAI expects negative free cash flow of $278bn between 2026 and 2030, according to a July presentation prepared for a computing deal and reported by the Financial Times. The same materials put compute and infrastructure spending at about $856bn across the period, against the roughly $600bn target the company gave investors publicly in February. The gap between the two numbers matters because it shows how much of the build is being financed by partners rather than by OpenAI itself.
What the July presentation contains
The document projects revenue of $350bn in 2030, compared with roughly $36bn this year. That is close to a tenfold increase in four years, and every other figure in the presentation is downstream of it. The burn forecast is not an independent estimate: it is what remains after the revenue assumption is subtracted from the spending assumption. A revenue miss therefore does not shave the burn, it compounds it.
The $278bn figure is an improvement on the roughly $305bn OpenAI projected in May for the same period, a difference of about $27bn. The compute and infrastructure line, meanwhile, has risen to about $856bn from the roughly $600bn target given publicly in February, roughly 43% higher. One caveat applies: the February figure was described as compute, while the July figure covers computing power and infrastructure, so the categories may not be identical and part of the gap could be definitional rather than real.
Spending can rise while burn falls when the spending does not sit on OpenAI's own balance sheet. OpenAI does not hold an investment-grade credit rating, which is why its financing runs through other companies. Nvidia has been in talks to guarantee $250bn of data centre debt, letting lenders price against the chipmaker's credit instead. 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 rather than OpenAI's.
What this means for businesses
For companies buying AI capacity, the practical consequence is that the infrastructure behind the models is being funded through vendor balance sheets, leases and guarantees rather than through a single rated borrower. That structure can keep capacity available and prices competitive in the near term, because partners have an incentive to fill the sites they have committed to. It also means the commercial terms a buyer sees may depend on which partner owns the asset, and a small company negotiating a contract has less visibility into that chain than a large one with its own procurement and legal review.
The limits are worth stating plainly. These are projections in a document prepared to win a computing deal, not audited accounts, and they have already been revised twice this year. OpenAI raised $122bn in March at an $852bn valuation and, according to the FT, is on track to exhaust that by 2028, two years before the projection period ends. The company has also paused its Stargate site in the UK over energy costs and copyright rules. A buyer should ask which entity holds the obligation, what happens to contracted capacity if a partner's financing terms change, and whether the revenue assumption behind the build is one the vendor can support with its own numbers.
The credit market has already shown where the strain sits: Oracle needed PIMCO to anchor $10bn of a $16.3bn data centre financing after US banks stepped back. When banks retreat from a name like Oracle, terms further down the chain are worse, and the guarantees and warrants are what the market required rather than a sign of strength. The marker to watch is whether the $856bn figure appears anywhere OpenAI can be held to it, such as a listing prospectus, since a number in a deal presentation and a number in a prospectus carry different consequences. The second marker is the guarantees: if partner balance sheets carry the difference between spending and burn, that exposure is theirs to report.
