The buildout outgrew the cash flow behind it

Oracle spent $1.74 for every dollar it earned

Last year Oracle put $1.74 into capital assets for every $1 its operations produced. CoreWeave put in $3.37. Both are now building faster than their own businesses can pay for, and the difference has to arrive as debt, as leases, or on somebody else’s balance sheet.

Neither number shows up in the figure everyone quotes. Across the five biggest spenders the ratio is a comfortable-looking 70%, and that average is doing the hiding. Big Debt Crises is blunt about why that is the normal case rather than bad luck:

Because debt bubbles typically emerge in one or a couple of markets, they are often hidden beneath the averages and can only be seen by doing pro forma financial stress tests of the significant areas.

So this is the ratio run name by name. Of the cash each company’s own operations generated, how much went straight back out as capital spending?

Every buyer moved up. The seller did not.

Those five spent $147 billion on capital assets in 2023 and $413 billion in 2025 — 2.8 times as much in two years, against operating cash flow that grew from $364 billion to only $588 billion. Spending outran the cash generating it. That gap has to be filled from somewhere, and where it gets filled from is the whole question.

Two names are already past the line

Oracle’s capital spending ran to 174% of its operating cash flow last year — a free cash outflow near $24 billion. CoreWeave’s ran to 337%. Both are now spending materially more on infrastructure than their operations bring in, which means the difference arrives as debt, as leases, or as someone else’s balance sheet.

Amazon has not crossed, but the direction is the same and the margin is thin: capex at 94% of operating cash flow, and a free cash cushion that fell from $32.2 billion to $7.7 billion in two years while revenue grew.

This is the change that matters, and it is worth being precise about what it is not. It is not that these companies are in trouble. It is that the buildout has changed category. Until recently it was paid for out of earnings, and a buildout financed from earnings has no credit cycle to be in. For two names, and soon for a third, that is no longer the arrangement.

The seller is on the other side of the same ratio

Nvidia converts about 94% of its operating cash into free cash, because it builds almost nothing: 6% of operating cash flow goes to capital assets, against $96.7 billion of free cash generated. Its 2025 revenue of $216 billion equals 52% of those five companies’ entire combined capital budget.

So one participant is accumulating the cash that the others are spending. And it has been putting a growing share of it back into them.

The number is in a filing rather than a press release. On 27 August, Nvidia disclosed $18 billion committed to equity investments for the remainder of its fiscal year, and $47.9 billion already held in private companies as of late July — roughly half a year of its own free cash flow, parked in the balance sheets of the ecosystem it sells to. Around that sit $3.5 billion of guarantees for companies leasing land and power, a standing promise to backstop young cloud providers, $2 billion into the power developer behind a Texas campus, and, this morning, $3.5 billion into MediaTek through convertible bonds — against a business MediaTek expects to produce about $2 billion of AI chip revenue this year.

That vendor financing is happening is not news, and this site has already made that argument through the CoreWeave capacity guarantee. What the filings add is the reason it exists. A supplier does not underwrite its customers while those customers can comfortably pay. The ratio above is what changed underneath the financing.

What a lender looks at, and what each of the three is doing

Dalio’s account of how a bubble inflates is not a mood or a warning. It is a description of a lending decision. A lender sizes a loan on three things:

lenders determine how much they can lend on the basis of the borrowers’ 1) projected income/cash flows to service the debt, 2) net worth/collateral (which rises as asset prices rise), and 3) their own capacities to lend. All of these rise together.

That last clause is the mechanism. The three inputs are supposed to be independent checks on each other, and in a bubble they stop being independent. Here is where each one currently stands.

The cash flows to service it. Capital spending at 70% of operating cash flow across the five, 174% and 337% at the two names past the line. The obligations are growing against cash flow that is not keeping pace.

The collateral. In the second quarter, Big Tech’s “other income” passed $160 billion, driven by marks on private AI stakes. For Amazon and Alphabet that was roughly $121 billion after tax — 66% and 71% of their quarterly profit. The collateral is being written up, and it is being written up on transactions priced inside the same circle.

The lender’s own capacity. This is the input that has changed most, and it is the one nobody had to think about before. The supplier has become a lender: $47.9 billion held in private companies, $18 billion more committed. Dalio’s phrasing for that stage is exact — “new types of lending institutions that are largely unregulated develop,” and “new types of lending vehicles are frequently invented.” Convertible bonds from a chip vendor, capacity guarantees from a chip vendor, and GPU-backed debt in special-purpose vehicles are what that sentence looks like when it is happening to you rather than in a chapter about 2006.

A majority of two of the world’s largest companies’ quarterly profit did not come from selling anything. It came from repricing what they own of each other. None of that is improper — marking equity stakes to market is required, not chosen. The narrower point is about earnings quality: profit sourced this way moves with valuations set inside the loop, and some of those valuations are supported by the financing the loop provides.

Graham asks the question this leaves open. In Security Analysis, whether a financed deal works means a price, a funding cost, a useful life, and an enforceable obligation from a customer who can pay. A mark-to-market gain on a supplier’s own customer answers none of those.

What would show this reading is wrong

The counter-evidence belongs in the argument rather than under it.

Investment is not consumption. Big Debt Crises is explicit that high debt growth funding consumption is the red flag, “since consumption doesn’t produce an income, while investment might.” Datacenters are investment. The test is not whether this is borrowed money; it is whether what the money builds earns enough to service it — and that is genuinely unresolved, not quietly settled against the buildout.

Vendor financing is ordinary. Equipment makers have funded buyers for a century. It only becomes a problem if the financed demand would not exist at an arm’s-length price, and nothing here demonstrates that.

The marks may be right. OpenAI and Anthropic have raised from unrelated investors at these valuations. A gain is not fictional because the holder is also a supplier.

Nothing here traces a completed circle. No filing cited shows a dollar leaving Nvidia and returning as Nvidia revenue. What the filings show is a ratio that moved, and who is on each side of it.

And the sign Dalio calls classic is not showing. He names one above all others: “when an increasing amount of money is being borrowed to make debt service payments, which of course compounds the borrowers’ indebtedness.” That is borrowing to pay interest, and nothing in these filings shows it. The money is still going into assets. A buildout financed on credit and a buildout refinancing its own interest are different objects, and this is still the first one.

Three things would change the reading, and all are visible in filings within two quarters: capital spending falling back under operating cash flow at Oracle and CoreWeave; other income shrinking as a share of Big Tech profit; or new borrowing at either name starting to cover debt service rather than construction.

Sources

Fundamentals are SEC EDGAR XBRL company facts — PaymentsToAcquirePropertyPlantAndEquipment over NetCashProvidedByUsedInOperatingActivities, with free cash flow derived as operating cash flow less capital spending. Values use the SEC’s calendar-year frames; Microsoft, Oracle and Nvidia have non-December fiscal years, so alignment across filers is approximate, and Nvidia’s 2025 column is the year ending January 2026.

Dated events are drawn from the coverage record and dated by first coverage rather than by deal close. The $18 billion committed and $47.9 billion held are Nvidia’s own disclosure of 27 August 2026; the other-income figures were reported on 31 August 2026.