The backstop is the disclosure

Jensen Huang announced Monday night that six firms — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR — will stand up independent financing platforms to move more than $500 billion of third-party capital into Nvidia AI factories. Ben Thompson wrote it up Tuesday morning. One number in the write-up does more work than the $500 billion: Nvidia is backstopping the deals itself, with up to 25 percent residual-value-based financing.

The backstop is the disclosure. Huang's pitch is that AI compute has become an investable infrastructure asset class, and every element of that pitch is a claim about residual value.

If institutional capital believed that, Nvidia would not be guaranteeing a quarter of it. Guarantees get priced. This one prices the distance between what Huang says a five-year-old GPU will be worth and what Apollo's underwriters were willing to sign.

The structure exists because the ordinary windows are narrowing. Oracle, Meta, Alphabet and Amazon raised a combined $108 billion in debt across all of 2025; by July 7 of this year they had raised $194 billion. Eighty-six percent of that issuance already trades at a higher yield than it priced at, and cover has fallen from 5x in February to under 2x. Microsoft is the last hyperscaler still funding capital expenditure out of free cash flow — $19.6 billion last quarter. When the bond market gets expensive, you go find capital that isn't in the bond market.

That capital has a name. Apollo owns Athene, KKR owns Global Atlantic, and both are annuity books: long-dated promises to retirees, funded by float that has to sit in something. Thompson names the exposure without flinching.

The person on the other end of that float bought a retirement product. He did not take a position on 2029 token prices.

The honest objection is that this is the job. Apollo and Brookfield underwrite long-lived productive assets for a living — toll roads, pipelines, fiber, generation capacity. They are not naive about residual value; they are the people who model it for a fee. That is the strongest thing you can say for the structure, and saying it isolates what makes this one different. A pipeline loses value to corrosion and to demand. A GPU loses value because Nvidia ships a better one, on a cadence Nvidia sets, at a price Nvidia sets.

Nvidia is guaranteeing residual value against a depreciation schedule Nvidia writes.

Huang's answer is fungibility and CUDA: the factories improve over time, so the asset holds. That is a bet that the software moat outruns the hardware roadmap, and Thompson's read is that the moat is already leaking — Anthropic hasn't depended on CUDA for years, and OpenAI is moving the same way on inference. If the frontier labs keep buying TPUs and Trainium because capital, not tokens per second, is the binding constraint, the redeployment story thins out at exactly the moment the financing depends on it.

Huang priced his own confidence at 75 cents on the dollar. The other quarter is sitting on a balance sheet that owes somebody an annuity.