Nvidia's $500bn GPU pool makes the seller the guarantor of its own demand

Nvidia said on the 10th that it had signed a memorandum of understanding with Apollo Global, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to assemble a dedicated $500 billion capital pool for AI infrastructure, with the six firms supplying funding, infrastructure expertise and support for the securitization process ^[1]. In some of those deals Nvidia guarantees up to 25% of the collateral value of the GPUs ^[2], which makes the manufacturer of the asset a partial credit backstop for the debt that buys it.

The load-bearing part is not the headline number, it is the ratings arbitrage. Bloomberg analyst Simon White, in research published on the 13th titled "Nvidia's Data Center Financialization Plan Raises Oversupply Risk," wrote that "by adding benefits to deal terms through the 25% guarantee, Nvidia has made it possible for senior debt to receive higher credit ratings" ^[3][4]. According to White, those higher ratings are what open the door for conservative lenders such as pension funds and large financial firms to participate ^[5]. Read plainly: Nvidia's own credit standing is being converted into cheaper funding for the customers who buy Nvidia's chips.

White compared the structure to AIG, which sold large volumes of credit default swaps on collateralized debt obligations containing subprime mortgages, reduced the capital banks had to hold, and was pushed to the brink of bankruptcy and rescued by the U.S. government when the loans defaulted ^[6][7]. He drew a distinction as well: "Nvidia's risk is not identical to AIG's. Nvidia is using its knowledge of its own product cycle to offset GPU depreciation risk and guarantee the recovery value of GPUs" ^[8]. That is a real difference. It is also an argument that Nvidia can forecast the resale value of its own products through a cycle it is itself accelerating.

The price of the debt tells you what the market thinks of the collateral without the guarantee. Ilya Kasyan, chief technology officer of GPU residual-value insurer American Compute, said on LinkedIn last month that GPU loan rates across the industry currently run from 9% to 18%, with company size the most important factor and long contracts from investment-grade borrowers at the low end ^[9][10]. The top of that range is double the bottom ^[1]. Nothing in an 18% coupon suggests a lender treating a GPU as durable infrastructure.

Utilization is the other number to hold onto. An April report from CAST AI put average GPU utilization at 5%, the waste arising because companies over-secure GPUs without sharing them or planning their use ^[11]. That leaves 95% of the fleet idle on average ^[2]. White's demand-side worry follows from this: as customers push AI model operators toward outcome-based pricing and fewer tokens, and as efficiency improves, rental prices could fall ^[12]. He dismissed the Jevons paradox rebuttal on the grounds that path dependency governs markets, and warned that demand for intelligence itself will eventually decline and computing costs will enter a prolonged slump ^[13][14]. He also noted that oversupply signals can be masked while capital keeps flowing into capex ^[15].

Watch the disclosure. If the 25% guarantee were applied across the whole $500 billion pool it would imply $125 billion of contingent exposure ^[3], so the question is how much of the pool carries it and whether it appears as a stated liability. Watch whether the senior tranches keep their ratings once GPU rental prices move, and whether the 9% to 18% spread narrows or widens as more paper is issued ^[9]. Google and other guarantors of GPU loans sit in the same position ^[16].