Oracle reports first-quarter fiscal 2027 results on Thursday, and nearly every preview written this week leads with the same number: $638 billion in remaining performance obligations, up 363% year over year. It is a genuinely startling figure, larger than the annual revenue of almost every company on earth, and it is doing an enormous amount of work in the bull case for a stock that has fallen close to 20% this year.
The part that keeps getting left out is Oracle’s own description of where it came from. In its record fourth-quarter and full-year results, the company disclosed that most of the RPO increase across the third and fourth quarters came from large-scale AI contracts in which customers either prepaid Oracle for GPU purchases or bought the GPUs themselves and supplied them to Oracle. Those two portions total roughly $75 billion. That is not cloud demand being booked in the ordinary sense. It is hardware financing and pass-through, sitting inside a number the market is valuing as though it were recurring infrastructure revenue.
Why the Composition Changes the Meaning
Remaining performance obligations is an accounting disclosure, not a quality judgment. It counts contracted revenue not yet recognized. A five-year database subscription and a customer wiring Oracle cash to go buy Nvidia hardware both land in the same bucket, and the bucket does not distinguish between them.
The distinction matters for three reasons that will all show up in Thursday’s numbers:
- Margin. Management has advertised an OCI margin profile in the 30% to 40% range. Revenue that is substantially the resale of customer-funded hardware does not carry software-like margin, and memory and GPU input costs have been climbing hard all year.
- Duration. A prepaid GPU contract converts to revenue as the hardware is delivered and the capacity is consumed. It is closer to a fulfilment schedule than to an annuity, which is what a 363% backlog jump implicitly promises.
- Credit. When customers prepay or supply their own hardware, they are extending Oracle working capital. That is a favourable arrangement for Oracle right now, and it is also a concentration of counterparty exposure to a small number of AI buyers whose own funding is not guaranteed.
None of this is hidden. Oracle put it in its own disclosures, which is exactly why the omission in the coverage is worth calling out rather than the company.
The Cash Statement Is the Honest Document
Strip out the backlog and look at what actually moved through the business last fiscal year. Oracle generated a record $32 billion in operating cash flow, up 54%, which is a real and impressive result. It then spent $55.7 billion on capital expenditure, roughly 83% of total revenue, and free cash flow came in at negative $23.7 billion. Guidance points to something near $70 billion of net capital outlay in fiscal 2027.
A company converting a genuine annuity backlog does not usually need to outspend its operating cash flow by that margin to do it. Oracle is building the capacity it has already sold, which is defensible, but it means the $638 billion is a claim on future construction as much as a claim on future customers. The financing gap has to be closed with debt or equity, and the market has already reacted once to that: Oracle’s shares sold off in June on a $40 billion capital raise even as the quarter beat.
This is the same structural question BusinessTech.News raised when CoreWeave’s $104 billion backlog turned out to be the collateral for its debt. Oracle is a vastly stronger credit than CoreWeave and the comparison is not about solvency. It is about a pattern: across the AI infrastructure buildout, contracted backlog has become the asset that justifies the borrowing that funds the capacity that services the backlog. The number is doing double duty.
What Thursday Actually Tests
Guidance for the quarter is cloud revenue growth of 58% to 64% and total revenue growth of 27% to 29%. OCI grew 93% to $5.8 billion in the most recent quarter while cloud applications rose 10% to $4.1 billion, a split that tells you how completely the story now rests on infrastructure.
The line to read first is not the revenue beat. It is whether Oracle breaks out, or is pushed to break out, how much of the RPO is prepaid and customer-supplied hardware, and whether that share is rising. If the $75 billion grows as a proportion of the total, the backlog is becoming less like a pipeline every quarter, not more.
Our Read
Oracle has not misled anyone. It disclosed the composition, and analysts who wanted the detail could find it. The failure here is in the reporting and the framing, including from outlets that should know better: a $638 billion headline number repeated without the $75 billion qualifier is not a summary, it is a distortion, and it has been repeated all week ahead of an earnings print that a lot of retail investors will trade.
Oracle should fix this itself, and it is easy to fix. Break out prepaid and customer-supplied hardware as a standing separate disclosure inside RPO, every quarter, the way any company reporting a mixed-quality backlog ought to. Investors can then decide what multiple to put on each piece. The company has an obvious incentive not to, because the undifferentiated number is more flattering, and that is precisely why the disclosure should be routine rather than discretionary.
The broader point is worth stating plainly. Backlog has quietly become the most important reported metric in AI infrastructure, and it is one of the least standardized. RPO rules give management wide latitude over what gets counted and when, and no two of these companies are counting the same way. Until that changes, treat every enormous backlog figure in this sector as a starting question rather than an answer, and ask what fraction of it is a customer buying compute versus a customer lending a vendor the money to build it.