A leaked OpenAI document projects $278 billion in cash burn through 2030 -- the same week the company is pitching investors a $1.2 trillion valuation

The Financial Times reported on September 18 that a confidential document OpenAI circulated in July -- shared as part of a major compute deal, not published for investors -- projects negative free cash flow of $278 billion between 2026 and the end of 2030. The same materials show revenue rising from $36 billion this year to $350 billion in 2030, and infrastructure spending of roughly $856 billion through 2030, the single largest line item in the model. At current spending rates, the document's own internal modeling shows the cash from OpenAI's March 2026 round -- $122 billion in committed capital -- exhausted by 2028.

What kind of document this actually is

The distinction that matters most here isn't any single figure -- it's what the document was for. As one detailed independent analysis of the leak puts it, "a projection circulated as part of a major computing deal is an underwriting document, not a forecast in the ordinary sense." It was shared with an infrastructure supplier deciding whether to commit multi-year capacity, which means it's built to answer a credit question -- can this company honor its commitments -- rather than to pitch growth to an investor. That reframes the $856 billion compute figure as the real constraint the document is actually about: it signals that securing capacity binds harder than selling the product does, since revenue climbs steeply in the same model while infrastructure spending is still the biggest expense on the page.

Two burn estimates, two different things

This isn't the first cash-burn number attached to OpenAI this year. Deutsche Bank had already estimated the company would run $143 billion in cumulative negative free cash flow between 2024 and 2029 before reaching profitability -- a figure the bank's own analysts called unprecedented, larger than Uber's roughly $18 billion in losses before its 2022 turn to positive cash flow, Tesla's $9 billion over nine years, or Amazon's roughly $1 billion over nearly five. Read next to the FT's $278 billion, that looks like the burn estimate doubled in a matter of months. It didn't -- the two numbers are answering different questions.

The overlap years (2026-2029) are the closest thing to an apples-to-apples comparison available, and even there the two sources aren't reconcilable from public information -- Deutsche Bank hasn't published its year-by-year model, and the leaked document's own annual breakdown hasn't surfaced beyond the 2026 and 2030 endpoints. What's clear either way: independent analysts and OpenAI's own internal materials now agree the company is burning cash at a scale with no real precedent among large technology companies, even if they disagree on the precise number by more than $100 billion.

What has to be true for both numbers to make sense at once

The document surfaced the same week Altman was reportedly in early talks with investors over a valuation between $1.2 trillion and $1.5 trillion -- up from March's $852 billion. Both things can be true without contradiction: a company can be worth more precisely because it's committing further ahead of demand it's confident will arrive, and a five-year, nearly-$300-billion cash gap is exactly what that kind of commitment looks like on a balance sheet before the demand shows up. What the leaked document adds isn't a reason to doubt the growth story -- OpenAI's own $36 billion-to-$350 billion revenue path implies real conviction -- it's a specific number for what covering the gap between committing and arriving actually costs.

Move the burn figure between the two reported estimates and the gap in what they imply is stark: at $278 billion, the March round alone covers roughly two years of the modeled average burn; at Deutsche Bank's $143 billion, it stretches to over four. Neither answer changes the more basic fact -- OpenAI has never operated near cash-flow break-even and isn't projected to for years -- but it changes how urgently the next round needs to close, and at what size, which is presumably part of why a document meant for a compute supplier is now shaping how investors read a funding pitch it was never written for.

"I think there are some players who are not managing [the risk of overinvesting] well or taking unwise risks." -- Dario Amodei, Anthropic CEO.

Amodei has an obvious rival's interest in that framing, and Anthropic's own reported cash burn is not zero either -- but the specific mechanism he's pointing at shows up directly in OpenAI's own numbers. Altman has separately acknowledged the company loses money on ChatGPT Pro subscriptions, and unlike Google, Meta or Amazon, OpenAI has no advertising, cloud, or retail business to offset AI infrastructure costs -- every dollar of the $856 billion has to be earned back through the same subscription and API revenue the $350 billion 2030 target is built on. That's the structural argument for why this projection reads as more architectural than alarming: the modeled 2028 cash exhaustion isn't a warning the company will fail, it's a description of continued access to capital markets being treated as an operating input, the same way compute and electricity already are.

None of these figures are reported results. $278 billion and $856 billion are unaudited projections from a confidential internal document, prepared to answer a specific counterparty's credit question and never meant for public reading. $143 billion is an outside analyst's model, not OpenAI's own number. $1.2-1.5 trillion is a valuation still being discussed, not one that has closed. What's real is the pattern underneath all three: a company simultaneously arguing, to a compute supplier, that its future cash flow can be trusted enough to extend credit against, and to investors, that its future revenue can be trusted enough to price at a 41% premium over March -- using numbers from the same five-year story, read by different audiences for different purposes.