Published · 11h agoLeadership2 min read
Jensen Huang says Nvidia may provide residual-value support of up to 25% on some financed AI projects. A published per-GPU-hour rate is what makes the other three quarters testable.
Context for builders, not their beat.See today for builders
Compiled by The Board RoomSomething wrong?How this is made
Residual-value support is a floor under the asset, not the loan. If a financing vehicle borrows against racks of GPUs and the chips turn out to be worth much less at the end of the term than lenders assumed, Nvidia's promise absorbs the first slice of that gap, up to a quarter of it on some projects [4]. Whatever falls below the floor stays with the vehicle's investors, and in these structures that means insurance floats and retirement capital placed by the same managers who arranged the debt [9][13].
So the resale value of a used accelerator becomes the load-bearing number in the chain, and nobody had to publish one. Ornn's index does not price a used chip. It prices the hour that chip rents for: a weighted average of verified compute transactions per GPU model, in dollars per GPU-hour, refreshed hourly, explicitly not list prices [1]. That earning rate is the input any residual model needs. Bank of America's Vivek Arya argued Nvidia's chips are unusually financeable because GPUs can be moved between operators and CUDA extends their useful life [8]. A marked rate per GPU-hour is what that argument looks like when someone has to put a figure on it.
The gap is duration. The open tier of the index is a trailing three-month daily window on five public GPUs, with hourly and realtime access gated behind keys [2]. Broadcom's parallel arrangement backstops five-year lease obligations [11]. Three months is five percent of sixty [15]. A lender underwriting the residual of a rack coming off lease in five years gets a spot rate and one quarter of history.
Ben Thompson of Stratechery reads the 25% as, in a certain sense, a price cut, with Nvidia spending its own margin to lower customers' cost of capital [6]. That is the accurate description. Goldman Sachs estimates AI-related financing already accounts for nearly a quarter of gross US investment-grade issuance, against AI investment approaching $600 billion this year [7]. The discount is aimed at buyers in a market that has already absorbed a great deal of debt, and the discount has a ceiling.
Ranked by verification strength, evidence, and original report placement.
Ornn's price index publishes an index_value that is the going rate of compute for a GPU model in US dollars per GPU-hour, a weighted average of all verified compute transactions across the market for that model, reflecting actual transactions rather than list prices; the GPU index is refreshed hourly, on the hour.
The keyless index tier is the trailing 3-month daily window on the five public GPUs; premium keys are daily-grain only, and hourly or realtime endpoints return 403 for them.
The index covers H100 SXM, H200, A100 SXM4, B200, RTX 5090 and RTX PRO 6000 WS, and currently tracks a single global region.
Nvidia CEO Jensen Huang said the company may provide residual-value support of up to 25% for some projects, effectively promising some protection against the possibility that the chips backing a financing are worth much less in the future than lenders expected.
Nvidia announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to create financing platforms intended to mobilize more than $500 billion for AI infrastructure, with the money largely coming from third-party investors and Nvidia's own risk limited and off balance sheet.
Ben Thompson, who writes Stratechery, calls the residual-value support "in a certain sense, a price cut": Nvidia using its own profits to reduce customers' cost of capital and make Nvidia-based data centers easier to finance.
The reporting this story was synthesized from, earliest first. Every link goes to the original.
