The Financial Times reported this morning that bankers for OpenAI and Anthropic are pressing the rating agencies to treat the two labs as investment-grade borrowers once they list. The reflex question is whether that helps Oracle, the supplier the tape reached for first. It is a reflex because the answer arrives before the thinking does — cheaper money for a customer sounds good for its supplier — and it is the wrong question, because it is a question about somebody else's balance sheet.
Rating agencies rate issuers. When S&P cut Oracle in July it was cutting Oracle: Oracle's capital spending, Oracle's cash flow, Oracle's leverage. A committee in New York deciding what OpenAI's paper is worth does not change a single term in a single Oracle credit agreement. The question worth asking is the borrower's: what has Oracle already promised to pay, whose rating is that promise priced against, and what would have to happen for the price to move.
The answer is in Oracle's own annual report for the year ended 31 May 2026, and the largest piece of it is not on the balance sheet at all.
Oracle disclosed $260 billion of additional lease commitments — "substantially all related to data center arrangements" — that had not commenced by 31 May 2026 and so were not reflected on its balance sheet. They are expected to begin between the first quarter of fiscal 2027 and fiscal 2029, for terms of fifteen to nineteen years. Against that sits $129.5 billion of borrowings that are on the balance sheet. The commitment Oracle has signed and not yet booked is twice the debt it has already booked.
The points
- $260 billion of lease commitments not yet commenced, substantially all data-centre, terms of fifteen to nineteen years, starting between fiscal 2027 and fiscal 2029. Disclosed in the annual report for the year ended 31 May 2026. One of those leases carries an Oracle guarantee of up to $3.3 billion of the lessor's own borrowing, maturing this month.
- $129.5 billion of notes payable and other borrowings on the balance sheet at the same date — $7.2 billion current, $122.3 billion non-current. Disclosed.
- $30.2 billion of operating and $7.7 billion of finance lease liabilities already on the balance sheet, separate from the $260 billion above. Disclosed.
- $13.3 billion of unconditional purchase obligations, primarily data-centre power arrangements, plus a further $19 billion of five-year cloud-infrastructure purchase commitments signed after the year ended. Disclosed.
- Free cash flow of −$23.7 billion in fiscal 2026, on capital expenditure that went from $21.2 billion to $55.7 billion in one year. Oracle expects "this upward trend to continue during fiscal 2027 and in the following fiscal years." Disclosed.
- BBB− since 9 July 2026, one notch above speculative grade, cut from BBB. The new rating carries a stable outlook. S&P's action, reported.
- No single customer accounted for 10% or more of total revenues in fiscal 2026, 2025 or 2024. Disclosed — and the reason the concentration argument has to be made about the book rather than about revenue.
Add the on-balance-sheet debt and leases to the commitments that are not on it and Oracle has promised roughly $441 billion against a contracted book of $638 billion. That total is our arithmetic on Oracle's disclosed figures, not a line Oracle publishes.
The FT story is about the labs' ratings, and it does not mention Oracle
Worth being precise about the news, because the tape has already fused two different things. The Financial Times piece, published today, is about the labs' own bankers pressing rating agencies to rate OpenAI and Anthropic investment grade after their IPOs, so the two can borrow from the very large pool of money whose mandates only buy investment-grade paper. It is a story about two loss-making private companies and the cost of their own future debt. Oracle does not appear in it.
The listing itself does reach other companies' books, and we have already argued one half of that: an Anthropic IPO reprices the stakes Amazon, Microsoft and Alphabet already carry. That is a marks event, and it lands on holders. A rating is a different consequence of the same filing, and it lands on borrowers. Oracle is a borrower — just not one that borrowed against OpenAI.
What has been circulating alongside it — that Oracle is refinancing debt tied to an OpenAI buildout, carrying a headline figure that appears in no Oracle filing — is a separate and older claim. Oracle's own financing announcement of 1 February 2026 lists OpenAI as one of six named cloud-infrastructure customers alongside AMD, Meta, Nvidia, TikTok and xAI. It does not name a customer-linked refinancing at all, and neither does the annual report.
Nothing happened at Oracle today. That is the point of writing this now rather than on the day something does: the mechanism is easier to see when the tape is not moving.
The two doors, and only one of them is Oracle's
There are two ways a credit story reaches a company's valuation, and the distinction is the whole piece.
The first door is the issuer's own cost of capital. It opens when the market re-prices the company's paper. It is the door S&P walked through on 9 July, and it is nailed shut against anything that happens to a customer's rating. Oracle's revolving credit agreement prices at SOFR plus a margin of 87.5 to 150.0 basis points "depending on the credit rating assigned to our long-term senior unsecured debt." Its term loan facility prices at 112.5 to 162.5 basis points on the same rating. Both grids read our rating. Neither reads OpenAI's.
The second door is whether the book converts. Oracle's $638 billion of remaining performance obligations is a promise from customers to pay, and its value depends on those customers being able to pay. A stronger OpenAI does make that marginally more likely. But that is not a rating effect travelling down a wire into Oracle's coupon — it is a slow change in the odds on a set of contracts, most of which sit years out. Oracle itself says only 12% of the $638 billion is expected to be recognised as revenue in the next twelve months, which is about $77 billion; 34% falls in months 13 to 36, another 34% in months 37 to 60.
The tape's version of today's story walks a customer's rating through the first door. It cannot get in.
Nvidia's exposure to a lab's credit is a guarantee. Oracle's is a lease
There is a supplier for whom a customer's rating genuinely does matter, and it is worth putting next to Oracle because the contrast is exact. On Nvidia's August call, Colette Kress qualified the balance-sheet support behind a quarter of next year's business by saying the "compute we ship will be consumed by investment-grade customers or those that are backed by one." Nvidia has written guarantees and taken equity. Those instruments reference the counterparty. If OpenAI is rated investment grade, something at Nvidia is genuinely released.
Oracle has written no such instrument on OpenAI. Its one disclosed guarantee of somebody else's borrowing, up to $3.3 billion, is of a lessor's debt — its landlord's, not its customer's. Oracle's exposure to the AI labs is not a promise it made about them. It is $260 billion of leases it signed in its own name, and a lease does not care who eventually fills the building.
That is why the same headline reads differently at the two companies. At Nvidia an investment-grade lab releases a contingent liability. At Oracle it releases nothing, because there was never a contingency — only a commitment.
What Oracle's own rating is actually wired to, in Oracle's own words
The strongest version of this argument is not our inference. It is Oracle's risk disclosure, which sets out what a downgrade does — and it is not mainly about coupons:
A downgrade could also reduce our access to, or increase the cost of, commercial paper or other short-term financing, affect the terms or availability of certain long-term commitments (including data center leases), limit eligibility to contract with certain customers, and increase collateral, letter of credit or other credit support requirements under certain contractual arrangements.
Read the middle two clauses against the $260 billion. Oracle is saying, in its own filing, that its rating can affect the terms or availability of the data-centre leases it is about to commence, and can limit which customers it is eligible to contract with. The supplier's credit gates the supplier's book. The customer's credit does not gate the supplier's rating.
It also has short-term paper to protect: the commercial paper programme was raised to $10.0 billion in March 2026, with $1.5 billion outstanding at year-end. Commercial paper is the financing that disappears first when a rating slips, and Oracle names it first.
A notch costs Oracle basis points. The leases cost it billions
Here is the size comparison that makes the first door look small.
Oracle's revolver is $10.0 billion, entered on 6 March 2026, and undrawn at 31 May 2026. Moving across the entire disclosed pricing grid — not one notch, the whole range from 87.5 to 150.0 basis points — is 62.5 basis points. Fully drawn, which it was not, that is $62.5 million a year. On the term loan facility, whose maximum commitment was $5.6 billion, the whole grid is 50 basis points, or $28 million at that size. Call the two together $90 million a year at maximum draw — roughly three cents a share, before tax. That arithmetic is ours, on Oracle's disclosed grids and share count.
Now the bond market, where the money actually came from. Oracle priced $25 billion of senior unsecured notes on 2 February 2026 across eight tranches, three-year to forty-year, at spreads of 95 to 195 basis points over Treasuries and coupons of 4.550% to 6.850%. The pricing sheet lists expected ratings of Baa2 (negative) / BBB (negative) / BBB (stable) and calls the deal, in the filing's own words, the "Investment Grade Notes Offering."
Those coupons are fixed. S&P's cut came five months later and does not touch a dollar of them. A downgrade does not reprice debt already sold; it reprices debt not yet sold. Oracle told the market on 10 June that in fiscal 2027 it "expects to raise approximately $40 billion through a combination of debt and equity financing including its previously announced $20 billion at-the-market equity issuance" — and told it on 1 February that it "does not expect to issue additional bonds during calendar year 2026 beyond this transaction." So the next bond is a 2027 event, and it prices at whatever Oracle is rated then.
Against all of that: $260 billion of leases, at fifteen to nineteen years, which Oracle owes whoever sits on the other side of the customer contracts.
What this does to our own numbers
Our Oracle model discounts at 9.5% and exits at 5.5x EV/revenue. Both are our assumptions, not disclosures, and the discount rate's stated reason is that 9.5% "pays for an investment-grade mega-cap carrying $129.5B of debt, negative free cash flow and about $40B a year of planned issuance." Six of our models justify a rate that way; Oracle is the one where the phrase is a single notch from being false.
On that basis, as of 23 August 2026, the model's base fair value is $221.32. Push the discount rate and it moves as you would expect:
| Discount rate | Fair value | Against base |
|---|---|---|
| 9.5% (ours) | $221.32 | — |
| 10.5% | $209.46 | −$11.86 |
| 11.5% | $198.23 | −$23.09 |
| 12.5% | $187.59 | −$33.73 |
Roughly $12 a share per hundred basis points. Every figure in that table is derived from our own assumption; none of it is a disclosure, and nobody can tell you how many basis points a notch is worth on an equity discount rate.
Now the other door. Our model carries a Concentration case, which asks what the book is worth if the largest customers stop renewing at current rates. It prints $133.13 — $88 a share below base. That is roughly two and a half times the effect of dragging the discount rate a full 300 basis points higher. Run both together and fair value is $110.48, roughly half the base. The second chart shows the four side by side.
The order matters more than the numbers. In our own framework the customer's credit is worth several times the supplier's rating — and it does not arrive through the rating at all. It arrives through whether contracts get served and renewed, which is a question about power, chips and delivery schedules, not about a rating committee. An investment-grade OpenAI would be a small, slow, welcome signal on the second door. It would do nothing to the first.
None of this changes the model. Our 9.5% was set for an investment-grade issuer on 23 August, and Oracle is still investment grade. The size of that "no" is the finding.
What to watch on 14 September
Oracle reports fiscal 2027 Q1, for the quarter ended 31 August 2026, on 14 September. Five things would settle something:
- How much of the $45–50 billion has actually been raised, split debt versus equity, and whether the $20 billion at-the-market equity programme has been used. The February plan was roughly half equity-linked; whether that half has been sold is the difference between a company funding itself with shares and one funding itself with paper priced off a BBB− rating.
- Whether the $260 billion of not-yet-commenced lease commitments grew, and how much of it commenced. Fiscal 2027 Q1 is the first window Oracle named for these leases starting. The moment they commence they become balance-sheet liabilities, and the leverage ratios the agencies quote change without Oracle signing anything new.
- The gross margin step-down. Management guided fiscal 2027 gross margin down "due to timing for the ramp up of our data center projects into their full revenue contribution plus impacts from mix." Gross margin was 70.19% in the quarter ended 31 May 2025 and 65.22% in the quarter ended 31 May 2026. How far it steps from there is the cleanest read on what this capacity actually earns.
- Whether the remaining performance obligation grows again, and whether Oracle says anything at all about its composition. It went from $138 billion to $638 billion in one year and Oracle does not break it out by customer. The half-is-OpenAI figure in circulation is S&P's estimate reported second-hand, not an Oracle disclosure, and it should be read as such.
- Whether Oracle repeats the "no additional bonds in calendar 2026" line. Breaking it would be the first hard evidence that the July downgrade changed the funding plan rather than just the headline.
Oracle's borrowings and lease liabilities, the $260 billion of lease commitments not yet commenced, the $13.3 billion of purchase obligations and the $19 billion signed after year-end, the remaining performance obligation and its recognition schedule, the revolver and term-loan pricing grids, the commercial paper programme and the downgrade risk factor are all disclosed in Oracle's annual report for the year ended 31 May 2026. Capital expenditure, free cash flow and gross margin are as Oracle reported them. The February bond terms are from Oracle's pricing term sheet of 2 February 2026. The S&P downgrade of 9 July 2026 and its estimate that roughly half the book is OpenAI are press-reported from that action, not Oracle disclosures; the Financial Times report on the labs' ratings is dated 8 September 2026. The $441 billion total and the basis-point and per-share arithmetic are ours. The 9.5% discount rate, the 5.5x exit and every fair value drawn from them are our assumptions as of 23 August 2026, on a $146.47 basis price that day — not forecasts, and not disclosures.
