Oracle's $664 Billion Backlog Puts the A.I. Buildout on Trial

A forty-year-old database company just signed up $209 billion of new, committed future revenue in a single year. The total — Oracle calls it remaining performance obligations, or RPO — now stands at $664 billion, about seven times the roughly $90 billion of revenue Oracle guides for the fiscal year ahead, and up 46% from a year ago.

That backlog is the whole story, and it is doing two jobs at once. It is the strongest piece of evidence that the artificial-intelligence buildout is real, and it is the strongest piece of evidence that the bill for it is large enough to break a balance sheet. Oracle's latest quarter makes both points in one document.

The backlog is the demand

RPO sounds like accounting. It is — but it is also the closest thing the industry has to a scoreboard. It is the portion of already-signed contracts that Oracle still has to deliver and get paid for: revenue that is promised, not yet on the books. A backlog that grows $209 billion in twelve months does not grow that way on enthusiasm. It grows because customers have signed papers.

The papers are coming from the frontier labs. A little over a year ago Oracle announced a roughly $300 billion cloud deal with OpenAI, and it has since been ramping capacity for OpenAI, xAI and Meta. The delivery is not theoretical: cloud-infrastructure revenue — the slice that runs other people's A.I. — jumped 121% in the quarter to $7.4 billion, and total revenue rose 30% to a record $19.3 billion, ahead of what Wall Street was expecting. Oracle said it brought online 850 megawatts of new data-center capacity and more than 300,000 graphics chips, and that those chips were about 98% utilized.

That last number is the one to fix on. Utilization near 98% is what separates "genuine demand" from "defensive inventory piling up in empty data centers." The capacity Oracle built last year is being used, not sitting idle. Whatever the long-run debate about overcapacity, the operating proof in this print is that, right now, demand is outrunning supply.

The bill is on Oracle

Here is the other half. To turn that backlog into delivered compute, Oracle is spending at a pace its cash flow has never seen. In the quarter it put $28.5 billion into infrastructure, and free cash flow — the money left after you build the business — came in negative $5.4 billion, even as operating cash flow hit a record. Full-year capital spending is guided to $90 to $95 billion, and it is funding much of the gap with new borrowing and a $20 billion stock offering.

The result is that a company rated investment grade for decades now sits one notch above junk. In July, S&P Global Ratings cut Oracle from BBB to BBB-, the lowest rung before junk status, citing weak cash flow — and, specifically, the concentration of that backlog in a single customer. That is the risk leg of the same print.

There is also a structural detail that makes the risk more than a timing question. Oracle says much of the newest backlog is either prepaid or "bring your own hardware," meaning the customer supplies the chips and Oracle's own outlay is lighter than the headline — and much of it will not reach revenue until fiscal 2028 or later. On its face that is reassuring. In practice it means the value of the backlog depends on whether those customers can keep paying, on schedule, through a multi-year ramp.

The $2 trillion question — and one customer

This is why the market keeps circling a number that is not Oracle's. Consulting firm Bain estimated, about a year ago, that by 2030 the A.I. industry will need roughly $2 trillion in annual revenue just to fund the compute it is building — and that, even counting A.I. savings, investors are still around $800 billion short. That is the real question, and it is a question about the frontier models and the labs that run them, not about Oracle's software.

Oracle is simply the most leveraged company in public markets standing in front of the answer. If the labs close that gap, a $664 billion backlog converts into a decade of compounding cloud revenue, and the debt looks like cheap fuel. If the gap does not close, this stops being a stock story and becomes a credit story — a database company that borrowed to build for customers whose own revenue has not caught up with their spending.

That is the binary the print narrows things down to, and it is sharpened by concentration. S&P named OpenAI a "key credit risk" behind the backlog. So the question is not whether A.I. demand exists — this quarter proves it. The question is whether enough of it comes from more than one lab, in time, from customers healthy enough to honor multi-year contracts.

A high-beta way to own the question

The long-term thesis is intact and, if anything, firmer than it was: the demand the strategy was built to capture is real, and it is arriving. I would not walk away from Oracle on this print.

But "intact" is not "attractive right now," and the two are different. The stock has fallen roughly 22% this year and about half from its highs, and it popped about 8% when the numbers landed — a partial reversal of a long decline, not a fresh entry on strength. What you own by holding Oracle is the highest-leverage pure play on whether that $800 billion revenue gap closes, funded by leverage that sits one notch from junk and concentrated in a customer S&P has flagged. That is a real position and a defensible one if you believe the frontier labs will generate the revenue to pay their bills — but it is a high-beta expression of a belief, not a cash-flow story, and the payoff is weighted to the back half, as the backlog converts over the next three years.

So the honest read of the print: it confirms the demand leg and exposes the financing leg, and it moves the decision from "is Oracle a good company" to the single, still-unresolvable question that sits behind every A.I. data center. Whether your capital is better deployed there, or with the customers or the chipmakers who carry less leverage to the same question, is a function of how confident you are in that one number — not of this quarter's scorecard.