Nvidia's $500B financing machine turns GPU depreciation into someone else's duration risk

The clause that will actually be priced here is the residual-value support, not the headline capacity. Jensen Huang said Nvidia may stand behind up to 25% of some projects, protecting lenders if the chips backing a financing turn out to be worth far less than underwritten ^[4]. Attach that rate to the whole program and Nvidia would be writing protection on $125 billion of used-GPU value ^[1]. It will not be the whole program, because "up to" and "some projects" are both doing work, and Nvidia has not disclosed the extent of any individual deal ^[3]. Ben Thompson reads the backstop as a price cut in a certain sense, Nvidia spending its own margin to lower customers' cost of capital ^[5].

The cushion exists because of who is on the other side. The paper ends up with insurers and pension funds that carry obligations running decades and therefore hunt for long-duration assets to match them, with private-credit managers doing the conversion work ^[9]. The collateral underneath is a chip whose depreciation schedule is the most contested number in the trade; the fear analysts have been voicing is a depleting pile of GPUs that swallows ever more capital ^[10]. Debt service comes from an AI tenant's lease or usage commitment rather than resale value ^[8], but the residual assumption is what determines recovery if the tenant stops paying. Bank of America's Vivek Arya argues GPUs are unusually financeable because they can be moved between operators and because CUDA stretches useful life ^[12]. That is a redeployment argument, and it holds only while there is a queue of buyers for last generation's silicon.

On size: more than $500 billion of intended mobilization ^[1] against AI investment approaching $600 billion this year ^[7] is roughly 0.83 of a single year at the current pace ^[2]. Which is the point. Goldman Sachs already puts AI-related financing at close to a quarter of all gross US investment-grade issuance ^[6], so the public bond channel is crowded. Insurance floats and retirement savings are the next pool, and Matt Levine's three-step version of the strategy says so without decoration: more private assets in ordinary retirement accounts, enormous private-credit and infrastructure funds, data centers built for AI to rent ^[14].

Morgan Stanley's Joseph Moore says the structure defuses the circular-financing worry, since third parties supply most of the money and Nvidia participates only to a limited degree ^[11]. On cash flows that is correct, and the stated design keeps the exposure off Nvidia's balance sheet ^[2]. But the residual guarantee means the vendor is insuring the future price of its own product while controlling supply and product cadence. Huang's framing, that AI factories can now be financed as productive infrastructure and that in AI compute is revenue ^[13], is the credit story a rating committee needs. What a note holder needs is narrower: where the 25% sits in the waterfall relative to the tenant's credit, and on what assumed useful life. Neither has been published ^[3].