Oracle's $664 Billion Backlog: The AI Number That Cuts Both Ways

Oracle's September 10 report opened with a number that worked. Revenue rose 30% to $19.3 billion, beating the $19.14 billion consensus, while cloud infrastructure revenue — the pure AI line — jumped 121% to $7.4 billion. And then the headline that pushed shares up about 8% in after-hours trading: the backlog of contracts not yet earned hit $664 billion, up $209 billion in a year and well ahead of the $639.9 billion analysts had modeled. Wall Street read it as proof that Oracle's all-in bet on AI data centers was paying off.

The interesting part is that the number is real. The debate is not whether Oracle has demand — it does.

The number that says Oracle got the supply side right

For two years the market's question about Oracle was whether it had missed the AI cloud. Then it did something the bigger three cloud providers mostly refused to do: it committed money and floor space at a pace that looked reckless before it looked smart. In the quarter just reported it delivered more than 300,000 GPUs to AI cloud customers — nearly triple what it delivered in the prior quarter — and brought 850 megawatts of new data-center capacity online. Management said customer demand for AI training and inferencing continues to outgrow supply, and it booked more than $30 billion of additional AI cloud contracts in the single quarter.

Because Oracle owns the capacity, it owns the biggest backlog in the industry — larger than Microsoft's. That is the supply-side win, and it shows in the segment that matters: cloud infrastructure at 121% growth is delivering the AI cycle to the income statement, not just to a slide.

That is the case for the number. Here is the part the same number carries with it.

The other half of the ledger is one customer and one date

Roughly half of Oracle's $664 billion backlog is attributable to a single customer: OpenAI. That concentration was the market's quiet worry through the year, and the quarter did nothing to reduce it.

Read the timing next. The biggest slug of that relationship is the roughly $300 billion, five-year compute deal announced a year ago, and it largely begins in fiscal year 2027 — meaning the largest contract in the backlog has not yet started generating most of its revenue. Stargate, the flagship facility built to serve it, is live in Abilene, Texas, but the revenue clock mostly starts next year.

Now put that demand against the cash it takes to build for it. Oracle has now reported negative free cash flow for five consecutive quarters — negative $23.7 billion for all of fiscal 2026 even as operating cash flow hit a record, because capital spending more than doubled. This quarter delivered a record $23 billion of operating cash flow against a negative $5 billion of free cash flow. The company guided fiscal 2027 net capital spending to between $70 billion and $95 billion, sold $20 billion of stock through an at-the-market program, and has said it will raise more. S&P Global cut Oracle's credit rating in July, citing exactly this cash-flow strain.

None of this means the demand is fake. It means the headline number is a dual signal: it is the strongest proof yet that Oracle owns the AI infrastructure buildout — and it is simultaneously a statement of leverage, one customer, and a conversion clock that mostly starts in 2027.

Why conversion, not the backlog, is the return driver

This is what separates Oracle the story from Oracle the stock. The backlog is a promise of roughly $664 billion over a long runway; management guides fiscal 2027 revenue to at least $90 billion, which is only about 12% of the current backlog converting in a year. If everything converts, this is a multi-year compounding machine. If AI capital spending slows, or OpenAI restructures the deal, or utilization disappoints, the backlog is just a very large number attached to a very large balance sheet.

The market has already been assigning value to this uncertainty for a year. Oracle is down more than 50% from its September 2025 peak, to about $151, and even after that fall it still trades near 35 times forward earnings and about six times trailing sales — a rich price on a company that does not yet convert backlog into free cash. So the question for a return is not whether the AI cycle is real. It is real for Oracle. The question is whether the conversion rate and timing — centered on one customer's 2027 start and a $70–95 billion annual build — justify the balance-sheet strain and the multiple that already assumes a lot of good news.

Demand was never Oracle's problem. Delivery, concentration, and the year 2027 are the numbers that actually decide what this backlog is worth from here.