Oracle's $664B Backlog Is Half OpenAI — Who Pays When AI Capex Peaks?

The number Oracle put in lights this week is $664 billion. That is the company's remaining performance obligations (RPO) — the dollar value of contracted future revenue it printed a record of after fiscal Q1 2027 results landed on September 10. Cloud infrastructure revenue up 121% to $7.4 billion, backlog surging on more than $30 billion of new AI bookings. The market has heard this song before: a giant backlog, an AI tailwind, and a stock that once rode it to a 36% single-day spike.

Here is the sentence the company would rather you not sit with. Of that $664 billion, roughly half is owed by OpenAI — a private company that has never made a profit and lost $20.9 billion in 2025. And only about 12% of the total backlog is scheduled to convert into actual revenue over the next twelve months, with another 34% in the twenty-four months after that. The rest of the promised cash sits beyond a three-year horizon.

The story works until this line. A backlog is not revenue, and revenue is not cash. The question is who pays, and with whose money.

The backlog that spends money before it earns it

Start with the mismatch between what Oracle promises and what it must lay out first. Oracle is not like most software companies, which book a contract and collect most of the money as they deliver. Building OpenAI's compute means building data centers and buying chips before the revenue arrives. In fiscal 2026 Oracle generated a record $32 billion of operating cash flow, then handed $55.7 billion back out the door in capital expenditures. Free cash flow for the year: negative $23.7 billion.

The gap is widening, not closing. Oracle raised its fiscal 2027 capex guidance to about $70 billion, plus another $20 billion to $25 billion in prepayments for capacity it has committed to buy. S&P Global Rates the buildout will push capital spending toward $90 billion to $95 billion and the free cash flow deficit to roughly negative $42 billion in fiscal 2027.

A deficit that large has exactly two possible funders: debt and equity. Oracle has already leaned hard on both. It carries more than $125 billion in debt and, in June, told investors it plans to raise about $40 billion during fiscal 2027 through a mix of debt and equity. S&P's research points to a roughly $20 billion equity component as part of that, on top of a $5 billion mandatory convertible preferred issued in February.

Here is the shareholder invoice in plain terms. Every dollar Oracle raises to build capacity OpenAI has promised to buy is dilution of the people who already own the stock — S&P estimates a $20 billion equity issuance would dilute existing shareholders by roughly 4.8%.

The unprofitable tenant

Now follow the same dollar to the other side of the lease. OpenAI's contract with Oracle is the largest single cloud deal ever signed: about $300 billion of computing over five years, beginning in 2027. It is the reason S&P Global Ratings cut Oracle to BBB- on July 9, 2026 — one notch above junk — citing OpenAI by name as a "key credit risk" and warning that OpenAI represents roughly half of Oracle's backlog. The rating agency also flagged a maturity mismatch: Oracle has locked into data-center leases running fifteen to nineteen years, against customer contracts closer to five years. If OpenAI stumbles, Oracle is left with long-term leases, prepaid chips, and no tenant.

The uncomfortable question is whether OpenAI can pay without someone feeding it money first. OpenAI has never generated a profit. In February 2026 it closed a $110 billion funding round — Amazon put in $50 billion, Nvidia and SoftBank $30 billion each — at a valuation S&P-era reports place near $840 billion. Its capacity to honor a $300 billion compute bill depends on continued rounds of exactly this kind of fundraising, and on the revenue from products that have not yet closed the gap to its $20.9 billion operating loss.

Test the circular hypothesis

This is where the case for "durable demand" and the case for "contingent claim" diverge, and it is worth being careful about the distinction. The circular-financing read works like this: the same pool of capital — SoftBank, Nvidia, Amazon, infrastructure funds — finances OpenAI's commitments on one side and Oracle's buildout on the other. Oracle raises debt and equity to build data centers. OpenAI signs a five-year contract to rent them. OpenAI's payment capacity is underwritten by the next funding round. The money travels in a loop, and a retraction anywhere in the loop stalls the whole system when the AI capex peak turns.

That is a hypothesis, not a fact. There is a benign version that fits the same numbers: OpenAI's own product revenue is genuinely compounding, and the $300 billion commitment is spread over five years precisely so it can be paid out of growing cash flow rather than new rounds. The 12%-in-twelve-months conversion profile, which looks alarming, is the mechanical result of a long, staged buildout — a five-year contract cannot recognize most of its revenue in year one no matter how healthy the buyer is.

The two readings are not distinguishable by the backlog number alone. They are distinguishable by documents Oracle has not yet shown us. Here is what would confirm or clear the risk:

• Conversion cadence. Oracle says about half the current backlog converts within thirty-six months. Watch whether each quarter's actual RPO-to-revenue conversion matches the disclosed 12%/34% ramp, and whether new bookings keep up with the dollars recognized. A backlog that converts on schedule while new contracts arrive is durable; a backlog that grows only because old contracts are re-booked longer is a claim being rolled forward.
• Prepayment recoverability and the cash bridge. Oracle's growing prepayments for capacity are cash out today against revenue years out. Oracle should disclose how much of that prepaid capacity is tied to OpenAI specifically, and how the negative free cash flow is being funded quarter by quarter.
• OpenAI's funding events and credit health. The single most decision-relevant document is OpenAI's next material fundraising or any change to its SoftBank-linked financing. S&P itself noted SoftBank reduced a share-backed loan to OpenAI. Each new round is evidence of payment capacity; a stalled round is evidence the tenant cannot cover the rent.
• OpenAI's own cash conversion. Watch whether OpenAI's product revenue growth outpaces its losses enough to cover its compute commitments from operations rather than from new equity.

Oracle's management, for its part, framed the 36-month conversion as a feature. But the strongest form of the defense — that this is contracted, staged, and back-weighted by design, not a fantasy — is also the form that concedes the risk: half of the contract's value is concentrated in one counterparty, and that counterparty's solvency is itself a function of the AI funding cycle.

The shareholder invoice

So what does a street-level investor actually own here? They own a company whose contracted future is real but slow — roughly $77 billion of it becomes revenue in the next twelve months, out of $664 billion promised — and whose current cash flow is funded by debt and by dilution of their own stake. The rating agency has already priced the concentration risk into a BBB- rating.

Here is the arithmetic of the downside. If OpenAI's payments stall and the contracts cascade — if Oracle cannot quickly find another tenant for compute built ahead of demand — the exposure is not the $300 billion headline. It is the portion of that capacity Oracle paid for before collecting, the board-hidden prepayments and multi-year leases, against revenue that never arrives. S&P's downgrade language is a formal, current statement that the agencies see stranded-asset risk as material, not a short-seller's rumor.

The case for hope is easy to price in: if OpenAI's product revenue compounds and the conversion ramp holds, Oracle converts a backlog larger than most countries' GDP into years of contracted cloud revenue, and today's cash-flow pain becomes tomorrow's free cash flow. The case for caution is equally mechanical: a backlog is only as good as the counterparty that owes it, and the counterparty that owes half of it issues a loss the size of Oracle's bad quarter while it funds itself from the very infrastructure boom that built this backlog.

For a buy-and-hold investor, the metric that matters most is not revenue growth and not the backlog headline. It is the conversion ramp and OpenAI's next funding fact, because those two numbers decide whether $664 billion is durable demand or a contingent claim waiting for the AI capex peak to stop the music. Sit with the 12%, because that is how much has actually earned its way onto the income statement so far. The other 88% is a promise, and the promise is only as strong as the person making it.