Oracle's order book reached $638 billion, up 363% in a year — and its free cash flow for the year was minus $23.7 billion

Every figure in this article is drawn from Oracle's fourth-quarter and full-year fiscal 2026 results, published 10 June 2026, and its SEC filings, or from CNBC and Data Center Dynamics as attributed. Oracle's fiscal year ends 31 May. Sections marked as analysis are identified as such.

Two numbers from Oracle's fiscal year, published in the same release. Remaining performance obligations — contracted revenue not yet delivered — reached $638 billion, up 363 per cent. Free cash flow was negative $23.7 billion.

Both are records. The first is the largest order book any enterprise technology company has ever reported; it is roughly nine and a half times Oracle's annual revenue. The second is the price of being in a position to report the first.

The year

Revenue for fiscal 2026 was $67.357 billion, up 17 per cent. Cloud revenue reached $33.989 billion, up 39 per cent, of which infrastructure — the part that competes with Amazon, Microsoft and Google — grew 77 per cent to $18.1 billion while the applications business grew 11 per cent to $15.9 billion. In the fourth quarter alone, infrastructure revenue grew 93 per cent.

Profitability held up: non-GAAP operating income of $28.926 billion on a 43 per cent margin, non-GAAP earnings of $7.63 a share, up 27 per cent. Operating cash flow rose 54 per cent to $32.0 billion.

It is worth separating the two Oracles inside those figures, because they are barely the same company. The applications business — enterprise resource planning, human resources, the software Oracle has sold for decades — grew 11 per cent at high margins and generates cash. The infrastructure business grew 77 per cent and consumes it. One is funding the other, and the ratio between them is changing fast enough that by next year the definition of what Oracle is will have changed with it.

And then capital expenditure of $55.663 billion took all of it and $23.7 billion more. Oracle raised $43 billion of debt during the year and has told investors to expect roughly $40 billion more of combined debt and equity in fiscal 2027, including a previously announced $20 billion equity issuance. Capital spending next year is guided at $90 to $95 billion gross, around $70 billion net.

Where $638 billion comes from

The order book is not diversified, and Oracle has not pretended otherwise. The single largest component is the five-year, roughly $300 billion cloud agreement with OpenAI announced in September 2025, covering some 4.5 gigawatts of compute. Co-chief executive Clay Magouyrk said the company signed $67 billion of AI infrastructure contracts in the fourth quarter alone.

The physical build is real and visible. The flagship Stargate campus at Abilene, Texas came online with more than a gigawatt of Oracle Cloud Infrastructure capacity during the year, with further sites under construction across several states and the programme's planned capacity expanding well beyond its original scope.

Concentration of this kind cuts against a principle Oracle spent forty years observing. Its historic strength was that no single customer mattered — tens of thousands of institutions paid maintenance every year regardless of the economy, which is why the company survived every downturn since 1980 without a bad quarter. A backlog dominated by a handful of AI counterparties is the precise inversion of that model.

Analysis: what a backlog is and is not

Remaining performance obligations are contracted revenue. They are not cash, they are not profit, and they are not certainty. They are a promise by a counterparty to buy something, recognised over the years in which it is delivered — and their value depends entirely on the counterparty's ability to pay.

That is the uncomfortable part of Oracle's position. Its largest customer is a private company that is itself burning extraordinary amounts of cash, whose revenue — while growing very fast — remains a fraction of what it has committed to spend on compute across all its suppliers. Oracle is building tens of billions of dollars of specialised infrastructure against contracts whose credit quality is unlike anything in its history of selling databases to banks and governments.

This is not a claim that OpenAI will fail to pay. It is an observation about the shape of the risk. Oracle has converted a diversified, prepaid, high-margin licence business into a concentrated, capital-intensive, debt-funded one, and the market has repriced the shares repeatedly in both directions as its confidence in that trade has moved.

The financing markets have noticed. Oracle raised $43 billion of debt in a single fiscal year, and reporting through the spring indicated that some lenders had grown more cautious about facilities tied to AI campuses where Oracle was the anchor tenant. Access to capital has not been a constraint so far. It is, however, now a variable in a business where it never used to be one.

Analysis: the case for doing it anyway

The case is straightforward and it is not weak. Oracle spent fifteen years as the fourth cloud, unable to win general-purpose workloads from three better-established rivals. AI training changed the basis of competition to one where Oracle could compete: raw availability of power, land, networking and GPUs, delivered fast. It moved earlier and more aggressively than Amazon or Google were willing to, and it won contracts none of them thought it could.

The 93 per cent infrastructure growth in the fourth quarter says the revenue is arriving, not merely promised. And Oracle's guidance for fiscal 2027 — $90 billion of revenue and non-GAAP earnings of $8.05 a share — implies that the backlog begins converting at scale this year, not in some indefinite future.

There is also a structural argument that gets less attention than it deserves. Data-centre capacity built for one AI customer is not stranded if that customer disappears; it is capacity, and demand for it currently exceeds supply by a wide margin. Oracle's downside is not a worthless asset. It is an expensively financed asset re-let at a worse price.

That argument holds only as long as AI compute stays scarce. The whole structure — the contracts, the debt, the depreciation schedules — assumes demand outruns supply for years. Every hyperscaler on earth is currently building against the same assumption, which is exactly the condition under which shortages historically end. Oracle's advantage is that its capacity is largely pre-sold; its exposure is that pre-sold at today's prices is a bet on tomorrow's.

Analysis: the number that decides it

Watch the cash flow statement rather than the backlog. Negative free cash flow of $23.7 billion is entirely defensible while capacity is being built ahead of revenue — that is what capital expenditure is for. It stops being defensible if the revenue arrives more slowly than the interest does.

Oracle now carries an obligation structure with three moving parts: what it has promised to deliver, what it has promised to spend, and what it has promised to repay. The first two are enormous and disclosed. The third grows every quarter the first two are out of step.

Larry Ellison spent four decades running a company that sold something once and collected maintenance forever. It has become, in about eighteen months, a builder of power-hungry industrial infrastructure financed on the bond market against a handful of very large customers. If the conversion happens on schedule it will be remembered as the boldest correct call in enterprise technology. The evidence is genuinely encouraging and genuinely incomplete.

For now, the fairest summary is the one the numbers give without interpretation. Oracle is growing its cloud infrastructure business faster than any company of its size, has more contracted future revenue than any enterprise software company has ever held, and is spending more than it earns to make good on it. All three statements are true simultaneously, and they will stay true for at least another year.