The dangerous moment isn’t simply when Big Tech has “too much debt.” Alphabet, Amazon, Microsoft and Meta can carry enormous amounts of debt because their underlying businesses still generate enormous cash flows.

The trap occurs when AI financing begins creating the appearance of AI demand that is then used to justify still more AI financing.

And there is fresh evidence that this deserves attention. Nvidia announced Monday it has lined up more than $500 billion in financing commitments from six of the largest capital allocators on earth — Apollo Global Management, Blackstone, BlackRock, Brookfield Asset Management, Goldman Sachs, and KKR — to fund AI infrastructure for its customers. The structure uses compute itself as collateral: special-purpose vehicles will issue tens of billions of dollars in debt at a time, backed by Nvidia hardware, with Nvidia reportedly prepared to provide residual-value support of up to 25% on individual projects. At the same time, combined Big Tech AI infrastructure spending is set to surpass roughly $730–750 billion this year.

Why this is a different kind of transaction than it looks like

On its face, this reads as good news — deep-pocketed, sophisticated institutions underwriting real infrastructure, diversifying the funding base away from hyperscaler balance sheets alone. Jensen Huang’s framing was explicit: “In AI, compute is revenue.” Strip the marketing language, and that sentence is a securitization pitch.

Toll roads, fiber networks, and power plants get financed at this scale because they throw off contracted cash flow for twenty years and nobody builds a materially better road next year. Nvidia is asking the same institutions that finance toll roads to treat a GPU the same way — while Nvidia’s own product roadmap is precisely the reason a GPU might not behave like one.

The mechanism of the trap, made concrete

Here is how financing demand becomes indistinguishable from economic demand, step by step:

1. Nvidia arranges financing that lets a customer acquire compute without paying upfront — the debt is serviced later, against the compute’s own expected earnings.
2. The customer’s purchase now counts as a sale in Nvidia’s revenue and as a capex commitment in the customer’s disclosures. Both figures are real, in the narrow accounting sense. Neither yet reflects a single dollar of end-user demand for what the compute will eventually produce.
3. That sale, that commitment, and the accompanying valuation of the financing platform become evidence — cited in earnings calls, analyst notes, and the next financing round — that AI demand is strong and accelerating.
4. That evidence lowers the cost of the next round of financing, because the market now treats compute as an established, revenue-generating asset class rather than a speculative one.
5. The next round of financing enables the next round of purchases, which produces the next round of “demand” evidence.

At no point in that loop does an end customer have to actually deploy the compute profitably. The loop can run entirely on financing mechanics, and every participant in it is behaving rationally given their own incentives — which is exactly what makes it dangerous rather than fraudulent. Critics have already raised concerns about circular financing, where a supplier funds the buyer who in turn funds the supplier, and have warned it can inflate the appearance of demand and valuation across an entire sector.

Why “compute as collateral” is the specific pressure point

Standard asset-backed lending works because the collateral has a defensible, independently observable value that doesn’t depend on the borrower’s own business succeeding. A mortgage-backed security is backed by a house; the house’s value is set by a broad market of buyers who have nothing to do with the original loan. Nvidia’s pitch depends on GPUs being similarly fungible — redeployable across customers if one operator fails, holding value independent of any single AI company’s success.

That