Deep Dive — OpenAI's cash gap shrank while its compute bill grew

The leaked number everyone is quoting is the smaller one, and it got better.

The Financial Times reported Friday that OpenAI expects negative free cash flow of $278 billion between 2026 and the end of 2030, according to a presentation the company prepared in July for a computing deal. The same materials put revenue at $350 billion in 2030 against roughly $36 billion this year, and compute and infrastructure at about $856 billion across the period. OpenAI declined to comment. The figures were confirmed as nonpublic by a person with knowledge of the deal who would not discuss them on the record.

Three numbers, one document, and the reporting picked the one that moves least. Here is the arithmetic the headline hides: an earlier projection in May put negative free cash flow for the same five-year window at roughly $305 billion. The July version is $278 billion. The figure being framed as an alarm is an improvement of about $27 billion on the company's own previous estimate, while the line that describes what OpenAI intends to buy went the other way — up from the roughly $600 billion compute target it gave investors publicly in February to $856 billion five months later, a gap of about 43 percent. Some of that is definitional: February's figure was described as compute, July's as computing power and infrastructure, so the categories may not be identical. But the direction of the revision is not in dispute.

Spending can rise while burn falls

The two movements are only compatible because the spending does not sit on OpenAI's own balance sheet. Free cash flow measures money leaving a specific legal entity. It is not a measure of obligations created, and the two diverge sharply when the build is financed by vendors, landlords and partners.

OpenAI does not hold an investment-grade credit rating, which is why its financing runs through other people's credit. Nvidia spent August negotiating the mechanics of a backstop on the company's Ohio campus and then cut a planned $250 billion guarantee to just under $120 billion after investors pushed back — a pullback we covered when it landed. The Ohio deal itself is the template: a 20-year lease at SoftBank's SB Energy campus in Pike County, roughly 8 to 10 gigawatts of capacity, with Nvidia guaranteeing a portion of the completed value and taking warrants in SB Energy that convert a capital commitment into an operating one paid for in equity rather than cash.

Oracle runs the same structure from the other side of the ledger. Its cloud infrastructure backlog reached $664 billion in remaining performance obligations, close to half of it attributable to a single OpenAI contract, while the company carried negative free cash flow of $5.4 billion and a $125 billion debt pile against a $90–95 billion capital expenditure plan. The data-center capex lands on Oracle's accounts, not OpenAI's. When Oracle needed $16.3 billion of data-center financing this year, US banks stepped back far enough that PIMCO had to anchor $10 billion of it — a useful read on the terms available further down the chain. That context matters for reading the $278 billion as good news. The obligations did not shrink; the entity that books them did.

The revenue curve is the actual forecast

Every other figure in the presentation is downstream of one assumption: revenue rising from about $36 billion to $350 billion in four years, close to a tenfold increase. The burn is not an independent forecast. It is what remains after the revenue assumption is subtracted from the spending assumption — which means a revenue miss does not shave the burn, it compounds it.

Set the curve against what the company has actually printed. OpenAI's second-quarter revenue grew 18 percent quarter over quarter to $6.7 billion, with its operating margin sinking further, per the Wall Street Journal. Annualized run rate has been reported above $40 billion, and the enterprise business has overtaken the consumer side by revenue. Anthropic, for comparison, more than doubled revenue to $11.6 billion in the same quarter and told investors its annualized run rate reached $65 billion by the end of July — the crossover we covered when Anthropic passed OpenAI on revenue for the first time. Two quarters of $6.7 billion annualizes to about $27 billion, not $36 billion, so the 2026 figure already assumes an acceleration through the back half. Getting from there to $350 billion requires roughly ten times that, sustained, for four years, in a market where Anthropic is spending to double its own compute and Chinese labs are shipping flagship models at a fraction of the price.

The spending side of that subtraction is the part already contracted. The revenue side is the part that has to show up.

Who is carrying the difference

This is the same structure we described when the Wall Street Journal tallied roughly $3 trillion in AI-related off-balance-sheet commitments across nine major technology companies against about $600 billion in reported capital expenditure — the gap between the size of the bet and what the books admit to, which we argued turns an investment program into a multi-year obligation the industry can no longer cleanly exit. OpenAI's presentation is a cleaner specimen of the mechanism because it is a single company's plan rather than an aggregate. Its compute bill rose 43 percent above the number it gave the public in February while its cash burn fell, and the reason is that the difference is being carried by counterparties who are not OpenAI.

Follow the exposure and it lands on companies with their own shareholders. Nvidia's quarterly free cash flow reached $48.5 billion, up roughly eighteenfold in three years, and it is spending that balance sheet and its credit rating on guarantees that let lenders price OpenAI's debt against a chipmaker instead of a startup. Oracle's leverage now trades on one customer's commitment. SB Energy's IPO filing disclosed it is "substantially dependent" on OpenAI. Those are not hidden instruments — they are in the filings — but they move the risk rather than removing it, and they make a demand miss ricochet across several balance sheets instead of landing in one.

The physical build has its own version of the same problem. The binding constraint on AI capacity stopped being chips and became megawatts, transformers and interconnection queues, the wall we mapped in the $1 trillion build-out that cash cannot fix. Committing $856 billion to computing power is a bet that the permits, the gas turbines and the skilled labor arrive on schedule. Gartner's revised forecast this weekend put worldwide AI spending at $2.7 trillion in 2026, up 49.5 percent, with infrastructure as the largest single category — so the aggregate demand the plan assumes is real and growing. It is also being counted the same way everyone else counts it.

The skeptic's case, fairly stated

The bear read has three parts, and two of them hold up.

First, the categories drifted. A public compute target of $600 billion in February became a private computing-and-infrastructure figure of $856 billion in July, and the company has revised its projections twice this year. These are documents prepared to win a deal, not audited accounts. A number in a pitch deck and a number in a listing prospectus carry different consequences, and only one of them has a prospectus on the horizon.

Second, the improvement is partly an accounting boundary rather than an operating result. If burn falls because a partner signs the lease and takes the warrants, the cost has not disappeared — it has been reclassified into someone else's capital structure, with equity dilution or a guarantee as the price. Read that way, $305 billion becoming $278 billion is not thrift; it is a financing decision.

Third, and weakest, is the Enron analogy that circulates around the off-balance-sheet pile. It does not fit. Enron hid its vehicles; these are disclosed in quarterly filings, and the US lease and consolidation rules were rewritten after 2001 precisely so a careful reader can reconstruct the exposure. Purchase obligations hit the books when goods arrive, unstarted leases roll on when the term begins. That is the standard, not a GPU-specific loophole.

Where the skeptics are right is on irreversibility, not concealment. A disclosed trap is still a trap. Once the leases are signed and the chip contracts placed, "we will slow capital expenditure next quarter" is a press release rather than a plan. The commitments do not evaporate on a demand miss.

The clock is the other story

OpenAI raised $122 billion in March at an $852 billion valuation, and the FT reports it is on track to exhaust that by 2028 — two years before the projection period ends. That is the context for the round the company has been discussing at a valuation above $1.2 trillion, talks that we noted when they surfaced last week. A company that needs capital before 2028 has a reason to be in the market before then, and Bloomberg reported the additional funding would give it room to push a listing back by a quarter or two, or to fund acquisitions.

Anthropic is on the same treadmill from a different position. It has told investors it expects about 5 gigawatts of available compute by the end of this year and to double that by the end of 2027, up from roughly 1.5 gigawatts last year, and it has slipped its own IPO marketing from October to after the US midterm elections. Both labs now need public or near-public capital to keep buying capacity at the rate their plans assume, which means the next twelve months are a financing story more than a capability story — regardless of what ships on the model calendar.

What to watch

Whether the $856 billion figure appears anywhere OpenAI can be held to it. A number in a deal presentation has no enforcement mechanism; the same number in a prospectus does, and that is the test of whether the plan is a forecast or a posture.

Watch the guarantees next. If partner balance sheets are carrying the gap between spending and burn, the exposure worth tracking is theirs — Nvidia's backstop, Oracle's backlog concentration, SB Energy's single-tenant dependency. And watch the revenue line, because it is the only input in the document that nobody has contracted yet.

If the spending is contracted and the revenue is not, which number should investors actually be pricing? Tell us in the comments.

Sources: Financial Times — OpenAI expects to burn $280bn by 2030 · The Next Web — $856bn compute vs the $600bn reset · NDTV Profit — OpenAI sees burning through $278 billion by 2030 · Wall Street Journal — why Big Tech's AI spending is $3 trillion higher than it seems · CNBC — Nvidia backing $105 billion in financing for OpenAI data center in Ohio · Gartner — worldwide AI spending to grow 49.5% in 2026