The $500bn compute asset class rests on a depreciation curve Nvidia once denied

Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR are working with Nvidia to assemble $500 billion in financing to turn compute into an investable asset class ^[1]. Anyone buying that paper is underwriting a residual value assumption, and Nvidia's chief executive said something incompatible with it about his own last-generation silicon a year ago ^[2][3].

Jensen Huang told CNBC that "this is really the first time that technology chips have become an investable asset class," calling them "revenue-generating assets" that are "productive, they're long-lived, they're fungible, they're flexible" ^[2]. At Nvidia's AI conference last year, selling the next architecture, he put it differently: "When Blackwell starts shipping in volume, you couldn't give Hoppers away," and "There are circumstances where Hopper is fine. Not many" ^[3]. The Verge, which reported both statements side by side, described the effect as whiplash ^[4].

The bridge between the two positions is software. In a post on X, Huang argued that Nvidia compute is not just a chip, that "their value is not fixed at installation," and that CUDA "continuously improves their output; the installed base remains productive well beyond its initial depreciation period" ^[5]. That is the load-bearing sentence in the whole structure. It makes the salvage value of the collateral contingent on the vendor choosing, year after year, to keep optimising for hardware it no longer sells.

For now the rent data supports the optimistic reading. The Verge reports that the price to rent older chips has been rising, and that Silicon Data projects it will keep rising through 2028 ^[6]. One cloud provider nearly doubled its price on Nvidia Blackwell B200 chips for a rental customer at contract renewal ^[7]. Scarcity pricing of that kind is good for current cash flows; it is not the same thing as a stable schedule an underwriter can amortise against.

Note also what "compute" is defined to include. Huang's description of the platform covers accelerated computing, networking, systems software, AI frameworks and a global developer ecosystem, with brick-and-mortar facilities absent from the list ^[8]. The Verge reads this as the GPU-backed loan rather than the full data center stack ^[9], and observes that most financing so far has gone to buildings: Blackstone has built a platform valued at $185 billion including facilities under construction ^[10]. The proposed compute programme is roughly 2.7 times that figure ^[11].

BlackRock chief executive Larry Fink told CNBC this is "the very beginning, like what it was when I started in the mortgage-backed securities market in the 1970s," and "a next future for financial engineering" ^[12]. Mark Rubinstein, a former hedge fund manager cited by The Verge, notes that mortgage-backed securities failed when mortgages were overproduced ^[13]. The parallel supply risks here are data center saturation and Chinese open-source models that need less compute while remaining fairly powerful ^[14], plus the unresolved question of whether frontier labs such as Anthropic and OpenAI can make money ^[15].

None of this is signed. The arrangement consists of memorandums of understanding ^[16], and Nvidia's $100 billion memorandum to invest in OpenAI last year did not happen ^[17].

Watch for the first actual issuance and the depreciation life and residual assumptions disclosed in it, and for whether Hopper rental rates hold once Blackwell supply is unconstrained.