Nvidia can now guarantee its own chips to unlock $500 billion in AI financing

Nvidia, led by CEO Jensen Huang, signed non-binding agreements on August 10, 2026 with six of the biggest names in finance to mobilize more than $500 billion in loans for AI data center buildouts. The six are Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR. The unusual part is what Nvidia is putting up to make those loans easier to underwrite: an optional guarantee, decided case by case, covering up to 25% of the residual value of its own chips if a financed project goes bad. That is a chip designer agreeing to eat losses on its own hardware so that banks and asset managers will lend against it. The announcement landed the same week Nvidia's own stock dropped on doubts about whether GPUs really hold value the way the pitch assumes.

how the financing actually works

The six firms are forming what the companies call compute financing platforms: pooled capital that lends directly to businesses building AI data centers, covering chips, servers, networking, and the buildings and power hookups around them. Nvidia's guarantee applies only to chips it signs off on for a given deal, and only up to a quarter of their assessed residual value, not the whole loan. Apollo's Jim Zelter puts the AI buildout's total eventual financing need at more than $8 trillion. Morgan Stanley separately estimates hyperscalers will spend $3.5 trillion on AI infrastructure between 2026 and 2028. Half a trillion dollars in new lending capacity is a start, not the whole bill.

Huang says a GPU should be financed like a power plant

Huang's argument is that GPUs behave less like laptops and more like infrastructure. Nvidia points to its own A100 chips, launched in 2020, as evidence: they are still being commercially rented six years later, which Huang says means a GPU's useful life runs closer to a decade than the three to five years standard accounting assumes. If that holds, a lender financing a data center full of Nvidia chips is financing something closer to a power plant than a fleet of laptops that are worthless in four years.

In AI, compute is revenue.

None of this is signed yet. The six agreements are non-binding memorandums of understanding, and the companies' own announcement notes they remain subject to final agreements, so $500 billion is a target, not committed capital. Investor Michael Burry has separately argued in public that GPUs depreciate faster than these financing plans assume, and puts a number on the gap: $176 billion in understated industry depreciation between 2026 and 2028.

Ben Thompson's Stratechery newsletter pushes the skepticism further. He argues the 25% guarantee works like a discount on the cost of capital that Nvidia is extending to its own customers, and compares it to Jay Cooke, the banker who began financing the Northern Pacific Railroad in 1870 by selling bonds directly to retail investors. That scheme collapsed a few years later and helped trigger a national financial panic. Thompson's point is that spreading risk into new corners of the financial system does not make the risk smaller. It just moves it somewhere less visible until it is not.

Why a build studio cares

This is not abstract for us. Every AI workflow we ship calls someone else's model over an API, and that API's price is downstream of exactly this kind of financing math. H100 rental prices went from $1.70 an hour in October 2025 to $2.35 an hour by March 2026, and current-generation B200 capacity rents for $5.30 to $7.05 an hour. Compute has been getting more expensive to rent, not less, before anyone has even found out whether a GPU is really worth financing for ten years. An agent build that assumes today's token price is a floor, not a ceiling, is planning around a fact that has not held for months.

Next step: read The Decoder's writeup of the financing structure next to Ben Thompson's skeptical breakdown on Stratechery. If you are building AI workflows on assumptions about where token prices go next year, we would rather have that conversation now: write to us at hello@gattyworks.com.