Oracle says customers have committed $664 billion. Its cash flow went the other way.

Oracle's first-quarter results for its 2027 financial year, reported on 10 September, contained one number that everybody repeated and one that fewer people did.

The first: remaining performance obligations of $664 billion, up about $209 billion in a year. The second: capital expenditure of $28.5 billion in the quarter, against free cash flow of minus $5.4 billion. Cloud infrastructure revenue grew 121% to $7.4 billion, and the company raised its full-year cloud infrastructure guide to at least $90 billion.

What "remaining performance obligations" actually means

RPO is the money customers have contracted to pay that the company has not yet delivered against. It is not revenue. It is not cash in the bank. It is not a forecast either — that is the useful part. A forecast is a company's opinion about demand; RPO is a stack of signed contracts, disclosed in a filing, that auditors have looked at.

So $664 billion is real in a specific, narrow way: customers have put their names to it. What it does not tell you is when it converts, at what margin, or what happens if a customer's circumstances change before delivery. A contract is a promise with a counterparty attached, and the value of the promise depends on the counterparty.

That is not an abstract caveat here. Our map records Oracle's announced compute arrangements with OpenAI, and the joint development of Stargate sites, as dated events rather than current operating relationships, precisely because that is what the evidence supports. A backlog concentrated in a small number of very large AI customers behaves differently from one spread across ten thousand enterprises.

The other half: someone has to build it

To deliver contracted capacity you must first buy it. The $28.5 billion of capital spending is the server racks, the GPUs, the buildings and the power. Free cash flow at minus $5.4 billion is what it looks like when you are spending ahead of delivery at scale.

The map shows the supply side of that spending as relationships rather than as a number: Oracle offering cloud compute built on NVIDIA hardware, buying GPUs from AMD, and leasing data-center capacity developed by Crusoe Energy. Every one of those is a line on the map with a source and a date behind it, and each is a place the $28.5 billion goes.

Management's answer to the cash-flow question is that most new AI contracts are prepaid or built on hardware the customer brings, which would narrow the gap between spending and receiving. That is a claim about contract structure made on an earnings call; it is not yet visible in the cash-flow statement.

Why the map colors this yellow and not green

Our rule is mechanical, which is the only way it stays honest: a company shows green when the recent, dated, sourced evidence about it points one way, and yellow when it points in more than one direction at once. Here, both directions are in the same press release. Record contracted demand is a tailwind for the argument that customers really will pay for AI capacity. Spending that far ahead of cash is a risk that sits on the same balance sheet.

Coloring it green would require ignoring half of what the company disclosed. Coloring it red would require ignoring the other half.

The question this leaves open

The central argument of the AI build-out, in one sentence, is: somebody has to pay for all of this. Oracle's backlog is the most concrete evidence anyone has produced that somebody has signed up to. Oracle's cash flow is the most concrete evidence of what it costs to get there first.

Both are true, both came from the same company on the same day, and the next four quarters are where we find out which one was the leading indicator.