Oracle

The first Oracle sells database software and business applications. When a bank records a transfer, an airline sells a seat or a government pays a pension, there is a good chance an Oracle database stores the record. Customers pay a licence and then, every year, a support fee of roughly a fifth of it, which they almost never stop paying because moving decades of critical data is risky, expensive and nobody's idea of a promotion. That business, together with Oracle's cloud applications (NetSuite, Fusion, healthcare software from Cerner), brings in roughly $39bn a year at very high margins.

The second Oracle rents computing power — Oracle Cloud Infrastructure, OCI — and has become one of the main builders of data centres for artificial intelligence. It grew 121% last quarter to $7.4bn. To serve that demand, Oracle expects to spend $90–95bn on capital projects in the year to May 2027 — more than its entire revenue — and most of it is being financed with borrowed money and newly issued shares.

A ten-year-old version: Oracle owns a toll road that everyone must use, and is using the tolls — plus a great deal of borrowed money — to build a giant new airport for one very large airline that has signed a long lease but has never yet made a profit.

★ Most companies on this board are one business. Oracle is two, and the market is pricing the argument between them. In September 2025 the shares hit a record when Oracle announced a $300bn computing contract with OpenAI and a backlog that had quadrupled in a year. Since then investors have looked at who has to pay for the buildings, and the shares have fallen by more than half — including, in the last week of June, the worst week for the stock since 2001. The business did not shrink. The question of who carries the risk got louder.

⚠️ A timing note. Oracle's fiscal year ends in May. Its second quarter of fiscal 2027 reports in mid-December, after this analysis. The last reported quarter ended 31 August 2026 and was published on 10 September.

Start with the good news, because it is real. In the quarter to 31 August, Oracle's operating cash flow was a record $23.1bn — much of it customers paying in advance. The company says most of the $30bn-plus of new AI contracts signed in the quarter were structured as prepayments or "bring your own hardware", so that they need no additional Oracle capital; the co-CEO Clay Magouyrk told analysts Oracle's own capital spending should now be seen as "decoupled" from how fast the business grows. GPU utilisation was 97.9%. Full-year revenue guidance was raised to at least $90bn, growth of roughly a third.

Now the rest of the ledger. Capital spending in the same quarter was $28.5bn — $18bn net of what customers funded — and free cash flow was −$5.4bn. For the full year Oracle guides to $90–95bn of capital spending, with no more than $70bn in net cash. S&P, which cut Oracle to BBB− on 9 July, projects free cash flow of about −$42bn this fiscal year. Borrowings have risen from around $104bn a year ago to about $125bn — $156bn including lease obligations — and in the quarter Oracle also sold $20bn of new shares through an at-the-market programme, after $5bn of mandatory convertible preferred and $25bn of bonds in February. Interest expense is up 55%.

★ Here is the whole case in one sentence: the order book is enormous, arrives slowly, and half of it is one customer — while the cost of serving it arrives now, and is borrowed. Only about 13% of the $664bn — some $86bn — is expected to become revenue in the next twelve months; 37% in the two years after that; half only after three years. The buildings, the chips and the interest have to be paid for long before most of it is recognised. That is a sound business model if every customer pays. It is a leveraged one if they do not.

Total revenue was $19.3bn, +30%; cloud revenue as a whole $11.6bn, +62%. Non-GAAP earnings were $1.92 a share. Three quarters of the growth is OCI. The software business that built Oracle's fortune is flat to slightly down — which is not a crisis, because it is being migrated to Oracle's own cloud rather than lost, but it means the entire growth story now rests on the capital-intensive half.

That matters for margins. A database licence costs almost nothing to deliver twice; a GPU cluster costs billions and depreciates in a few years. Management itself has said the non-GAAP operating margin — 42% last quarter — will keep drifting lower as OCI becomes a larger share of the mix.

Two days ago, in our follow-up on Microsoft, we found that OpenAI supplies at least a third of Microsoft's $678bn commercial backlog. Here it is about half of Oracle's $664bn. Between the two companies, more than $500 billion of contracted future revenue depends on the ability of a single company — one that has never reported a profit — to keep raising money and paying its bills.

We do not predict that OpenAI will fail to pay. It has extraordinary revenue growth, deep-pocketed backers and every incentive to honour contracts on which its own future depends. But concentration is not a prediction; it is a shape. Microsoft carries its third on a AAA balance sheet with $183bn of operating cash flow. Oracle carries its half on a BBB− balance sheet with negative free cash flow, $156bn of debt and a dividend it is borrowing to pay. The same customer is a manageable risk for one and the central risk for the other.

★ The asymmetry that decides it: if OpenAI pays, Oracle's backlog becomes one of the great revenue streams in technology and today's price looks cheap. If OpenAI renegotiates — as large customers do when their finances tighten — Oracle is left with specialised buildings, fast-depreciating chips and the debt that bought them. The first outcome is shared with every other supplier. The second falls hardest on the most indebted one.

Type: switching costs — in the old business. We score the moat 7: the database franchise alone would earn a 9, but the capital, the growth and now the risk have moved to a business where Oracle's advantage is execution and appetite rather than something competitors cannot copy.

Integrity: no veto, but a live question. Two lawsuits allege that Oracle understated how much it would need to borrow (Part X). We make no finding on them; we note that the pace of financing announcements — $18bn of bonds, then $38bn of loans eight weeks later, then a $45–50bn plan — surprised investors each time.

Capital allocation: the sharpest reversal on our board. In fiscal 2021 and 2022 Oracle bought back $39bn of its own shares, shrinking its equity almost to zero. In the quarter just reported it sold $20bn of new shares, at prices far below where it bought them. Diluted shares rose from 2,909m to 3,000m in a year. A company that shrank its share count for a decade is now growing it to pay for a single strategic bet. That may prove brilliant. It is not what an owner signed up for, and it earns a 4.

★ Nobody should own Oracle for its dividend today. It is small, it is not growing, and it is not covered by the cash the business generates. Compare Microsoft, whose dividend is covered 2.3 times by free cash flow even at the peak of its own AI spending. The difference is not the strategy; it is the balance sheet that carries it.

So is Oracle cheap? On earnings, arguably yes: under 17 times next year's consensus, falling to single digits by 2029 if the backlog converts. On cash, it is not measurable, because there is none. On the two-Oracles arithmetic, the price asks you to pay a modest premium over cost for a leveraged AI landlord whose largest tenant has never made a profit.

★ This is not an overpriced company. It is an unpriceable one. The outcome is binary in a way most of the board is not: if OpenAI pays, $137 will look like a gift; if it does not, the debt decides what the equity is worth. A Buffett-style investor does not need to know which — only that he cannot know, and that the cost of being wrong is carried by the lender first and the shareholder second.