Oracle's $664 Billion AI Backlog, Read the Way a Credit Analyst Would

Oracle reported $664 billion of remaining performance obligations at 31 August 2026, and its own filing says only about 13% of it becomes revenue in the next 12 months. Oracle RPO is the largest contracted backlog in enterprise software. Roughly half of it is reported to sit with one customer that does not yet earn enough to pay for it, and Oracle has to build the data centres before it can send the first invoice.

This post is written for a founder or director. You either buy AI infrastructure on multi-year terms or sell into a market priced off backlogs like this one. The useful output is therefore a reading method you can reuse, not a view on Oracle's share price.

What Oracle RPO actually measures, and what it does not

Remaining performance obligations is an accounting disclosure required under ASC 606, the US revenue recognition standard. It is the total value of contracts signed for which the vendor has not yet delivered the goods or services. It includes deferred revenue already billed and contracted amounts not yet billed.

That definition makes RPO more disciplined than a sales pipeline. A pipeline is hope. RPO requires a signed contract with enforceable terms. That is why investors treat it as a forward revenue signal. It is also why Oracle's stock rose 36% in a single day in September 2025. That was when RPO first jumped to $455 billion, up 359% from a year earlier, per Fortune's report at the time.

RPO is a contract number, not a cash number

A lender cares about cash that arrives, not revenue that is promised. RPO tells you neither when cash arrives nor whether it will. It records what a counterparty has agreed to pay, on a schedule that depends on the vendor delivering first.

Fortune put the caveat plainly in 2025: remaining performance obligations "are not guaranteed revenue; customers can delay, renegotiate, or even cancel portions". The legal enforceability of any single contract depends on terms nobody outside the two parties has seen.

An equity investor reads a backlog for upside. A credit analyst reads it for the ways it fails to pay. The second lens asks three questions the first one skips: who owes the money, when it is owed, and what the vendor must spend before it can collect.

My position is that the credit lens is the correct default for any AI infrastructure backlog. The upside is already in the headline number. The risk is in the footnotes, and that is where the rest of this post stays.

The Oracle OpenAI contract and the concentration problem

Oracle signed a compute agreement with OpenAI in mid-2025, widely reported at about $300 billion over five years. That single deal accounts for most of the jump that made Oracle's backlog famous. A year later, the share is still large.

ERP Today's coverage of the fiscal 2027 first quarter states that roughly half of the $664 billion backlog is tied to OpenAI. S&P's July 2026 rating action described OpenAI as roughly half of the earlier $638 billion figure, according to PPC Land's account of the downgrade.

The disclosure that answers the wrong question

Oracle's fiscal 2026 10-K states that no single customer accounted for 10% or more of total revenues in fiscal 2026, 2025 or 2024. That sentence is true and almost irrelevant. It measures revenue already recognised. The OpenAI capacity mostly has not been delivered, so it has barely reached the income statement.

A credit analyst would want concentration measured against the backlog, not against trailing revenue. Oracle does not publish that split. Every concentration figure you will read, including the ones in this post, is a third-party estimate. That is the first gap in the record, and it is a big one.

The same 10-K does name the risk without naming the customer. It says Oracle enters "large, long-term customer cloud arrangements that require us to make significant infrastructure investments". It adds that returns depend on "the ability of our key customers to meet their contractual obligations". That is as close as the filing gets.

Can the counterparty pay?

A $300 billion contract over five years works out to an average of about $60 billion a year, which is my arithmetic on the reported figures rather than a disclosed schedule. The actual payment profile is not public and is likely to ramp.

Set that against what OpenAI earns. The Financial Times, as summarised by TNW in September 2026, reported an OpenAI presentation projecting about $36 billion of revenue in 2026. It also projected negative free cash flow of $278 billion between 2026 and the end of 2030. The same presentation put compute and infrastructure commitments near $856 billion.

None of that means OpenAI will default. It means OpenAI's ability to pay Oracle depends on raising capital and growing revenue roughly tenfold by 2030, which the same projection targets at $350 billion. A lender would call this a contract backed by a growth plan, not by cash flow. The structure behind it is covered in the breakdown of how the AI circular deals actually work, and the valuation underpinning OpenAI's fundraising in the analysis of OpenAI's run-rate multiple.

When the Oracle backlog becomes revenue: the 10-Q schedule

This is the most useful sentence in Oracle's quarterly filing, and it is rarely quoted in full. The 10-Q for the quarter ended 31 August 2026 sets out the schedule. It expects to recognise "approximately 13% as revenues over the next twelve months, 37% over the subsequent month 13 to month 36, 34% over the subsequent month 37 to month 60 and the remainder thereafter".

Percentages are Oracle's own disclosure. Dollar amounts are Zan Digital arithmetic on those percentages and are rounded. Oracle states the final bucket only as "the remainder thereafter".

Read the movement, not the total

Between May and August, the near-term buckets grew and the far bucket shrank. The next-12-month share went from 12% to 13%. The beyond-five-years share fell from 20% to 16%. That is what a backlog looks like when contracts start ageing toward delivery rather than piling up at the back.

This is genuinely good news for Oracle and it deserves to be said. A backlog that only grows at the far end is a warning. One that migrates forward is converting. I would watch this single ratio each quarter before I looked at the headline.

Oracle raised its fiscal 2027 revenue guidance to at least $90 billion, according to ERP Today. The roughly $86 billion due within 12 months covers most of that, because RPO also includes existing software support and cloud subscriptions, not only AI capacity. That cross-check is useful in reverse: it tells you the AI-specific portion of near-term conversion is smaller than the headline suggests.

The capex Oracle must spend before it can bill

A software subscription backlog costs little to deliver. An AI compute backlog is the opposite. Oracle has to acquire land, power, buildings and accelerators before the customer pays for the first hour of capacity. The order of operations is spend first, bill later.

The 10-Q shows $28.5 billion of capital expenditure in a single quarter, against $8.5 billion a year earlier. For the whole of fiscal 2026 the figure was $55.7 billion. Guidance for fiscal 2027 is unchanged at $90 billion to $95 billion, per ERP Today.

The lease commitments that sit outside the debt figure

Capex is only the owned part of the build. Note 6 of the 10-Q discloses $288 billion of additional lease commitments, substantially all data centre arrangements. They are expected to commence between the second quarter of fiscal 2027 and fiscal 2029, for terms of 15 to 19 years.

That figure rose $28 billion in one quarter. These leases do not appear on the balance sheet until they commence. The wider pattern across the hyperscalers is measured in the breakdown of data centre lease commitments still off the balance sheet.

Note the mismatch a lender would flag. The leases run 15 to 19 years. The flagship customer contract is reported at five. Somebody has to fill those buildings for a decade after the original deal ends, and nothing in the backlog yet guarantees who.

How the build is funded: debt, equity and prepayments

The funding source tells you how confident the vendor and its lenders are. Oracle's has changed three times in a year, and the current mix is worth reading closely.

The fiscal 2026 10-K records $43.0 billion of senior notes issued that year. Total notes payable and borrowings were $129.5 billion at 31 May 2026, up from $92.6 billion a year earlier. By 31 August the balance was about $125.3 billion, because Oracle repaid $4.2 billion and issued no new bonds in the quarter.

That pause was not a choice made in a vacuum. On 9 July 2026, S&P cut Oracle's long-term rating from BBB to BBB-, one notch above speculative grade. PPC Land quotes the agency: "we now recognize that that we underestimated the scale of the investments required". S&P said a further cut could follow if adjusted leverage stays above 4.5 times. Debt became expensive at the margin, and interest expense rose 55% year on year to $1.4 billion in the quarter, per ERP Today.

Equity and the customer's own cash

The 10-Q shows $19.9 billion raised through an at-the-market stock programme in the quarter, with shares outstanding rising from 2,880 million to 3,024 million. That is about 5% dilution in three months.

The quarter also brought $11.4 billion of customer prepayments, against $4.6 billion for the whole previous year. ERP Today reports management describing most of the more than $30 billion of new AI contracts as prepayment or bring-your-own-hardware structures, which shift capital onto the customer.

I think the prepayment shift is the most important change in Oracle's model this year, and it cuts both ways. It reduces Oracle's capital at risk. It also inflates operating cash flow: ERP Today reports operating cash flow of $23 billion and free cash flow of negative $5.4 billion, with cash capex of $18 billion net of prepayments. Prepaid cash spent on a build is still spent.

The cash side of Oracle's story, including how it shaped the company's 2026 job cuts, is set out in the analysis of the Oracle layoffs and the cash flow hole. A comparable debt structure at a smaller neocloud is examined in the CoreWeave debt stack stress test.

Where this argument is weakest

A post that reads Oracle's backlog only for risk would be making the same error as one that reads it only for upside. Here is the case against my own framing.

The utilisation and pricing evidence is strong

ERP Today reports Oracle citing 97.9% GPU utilisation and renewal capacity resold at a 20% premium. If those figures hold, the capacity Oracle is building has buyers beyond OpenAI. ERP Today also reports that the non-OpenAI backlog more than doubled over the past year. A diversified half of $664 billion is still about $330 billion, which is larger than most software companies' entire market value.

Prepayment structures genuinely change the risk

A customer that pays upfront or supplies its own hardware has absorbed the very risk this post worries about. If most new contracts really run that way, the gap between backlog and cash narrows for each new deal. My criticism is that the older, larger contracts were signed before this structure existed, and nobody outside Oracle can tell which bucket they sit in.

The figures here are partly secondary

The concentration estimate, the cash flow totals and the guidance all come from trade press or a summary of a rating action. None are documents I could open directly. Oracle's own earnings release and S&P's full report were not accessible to me. The 10-Q figures are primary. Treat the rest as Tier 2, and check them against Oracle's investor day in October, which ERP Today says should give the first detailed view of how the backlog converts.

A five-test method for judging any AI vendor's backlog

Oracle is the clearest case, but the same questions apply to any AI vendor that announces a large contracted backlog. That includes neoclouds, chip suppliers with take-or-pay agreements, and software vendors reporting multi-year AI commitments.

I would not average them. Tests 1 and 4 combine into a single question: if the largest customer stopped paying, what share of the backlog survives? For Oracle that answer is about half, which is still enormous, so the right conclusion is "large but lumpy" rather than "illusory". I disagree with commentary that treats the OpenAI dependence as proof the backlog is fake. A contract with a stretched counterparty is a credit risk, not a fiction.

Test 2 is the one that moves fastest and costs nothing to track. Test 3 is the one that decides whether the vendor survives a delay. A vendor with heavy pre-delivery spend and a slow conversion schedule is exposed to every month the customer's plans slip. The accounting choices that make that exposure look smaller on the income statement are covered in the analysis of GPU depreciation and AI earnings.

Frequently asked questions

Where to start

If you buy AI infrastructure on multi-year terms, open your largest compute contract this week and find the clause that covers vendor capacity delays. Then ask your vendor one question: what share of its own backlog sits with its largest customer? A vendor that will not answer has told you something about concentration.

If you sell into this market or price off it, put one number in your calendar for each Oracle quarter: the next-12-month RPO percentage. If it keeps rising, the backlog is converting. If it stalls while capex keeps climbing, the build is running ahead of the customers, and you will see it there before you see it anywhere else.

Related reading