Nvidia put its name to memorandums of understanding with six of the world’s largest money managers, who will build lending platforms for the customers who buy its chips. CNBC reported Monday that the group, made up of Apollo Global Management, BlackRock, Blackstone, and Brookfield Asset Management, alongside Goldman Sachs and KKR, aims to put upward of $500 billion of outside capital behind data centers and hardware purchases. Hyperscalers, frontier labs, and enterprises would get their capacity without funding it from their own balance sheets.
That last clause is the product. Nvidia is not selling a chip here. It is selling an underwriting assumption.
The assumption is that a used accelerator is collateral. Jensen Huang told CNBC that his silicon now qualifies as an investable asset class, which he framed as a first for the semiconductor industry, and described the hardware as productive, long-lived, fungible, and flexible. Strip the adjectives and one claim carries the weight: a lender who forecloses on a rack of GPUs can resell them to somebody else at a price it can forecast today. Every dollar of the $500 billion prices off that.
CNBC’s account does not explain how the lenders got comfortable with it. The story describes memorandums, which bind nobody, and quotes principals rather than terms. It does not say who absorbs the loss if resale prices fall, whether Nvidia extends any credit support or repurchase commitment, what advance rates apply, or what useful life the models assume. That silence is the most consequential thing in the piece. If the depreciation risk sits entirely with the funds, this is a genuinely new asset class. If any of it routes back to the vendor, it is vendor financing with more counterparties, and the balance sheet absorbing the shortfall belongs to the company selling the chips.
Larry Fink reached for the precedent himself. BlackRock’s chief executive told CNBC he views the project as the opening of a new era in financial engineering, and the parallel he chose was the moment Wall Street first bundled home loans into tradable paper, roughly five decades ago. He offered that as praise. Fink added that some money has already been gathered and that his firm will raise considerably more. Securitization works when the collateral behaves the way the model says it behaves, and it fails in precisely one way when it does not.
Jon Gray of Blackstone extended the housing comparison, saying AI compute will be treated as financeable in the way mortgage lenders treat homes. Houses are a flattering comparison for a reason nobody raised on air. A house depreciates slowly and is not superseded by a faster house on a published roadmap. Gray also said AI usage across Blackstone portfolio companies rose sevenfold this year, which is the demand case holding the whole structure up.
The timing is not incidental. The announcement follows a July selloff in which investors began questioning the payback on Big Tech’s AI spending, and it follows a Moody’s warning that heavy capital expenditure is compressing free cash flow at the hyperscalers and pushing them toward larger debt loads. Routing GPU purchases through third-party credit vehicles answers that pressure directly. It does not shrink the spending. It changes who reports it and who is exposed when the assets underperform.
Watch the collateral assumption rather than the headline number. Tom’s Hardware reported Monday, citing The Information, that Nvidia has been testing Rubin Ultra configurations with as little as 192 GB of memory and with HBM4 substituted for the HBM4E it originally announced, in response to memory supply constraints. Nvidia told that outlet its roadmap is intact. The specification of a coming generation is still moving while it is being built, and residual value curves for today’s accelerators are drawn against assumptions about that same roadmap. Anyone modeling a smooth five-year depreciation schedule is modeling a product cadence that is not settled.
Apollo and Blackstone have already arranged debt and equity for AI companies including Anthropic, so the capability is real. What is untested is the repossession scenario. Nobody has yet liquidated a large fleet of prior-generation training GPUs into a market that no longer wants them, which means the recovery assumption underneath $500 billion of lending has no historical print behind it.
For anyone signing a compute lease or a GPU-backed facility in the next two quarters, the term worth negotiating is not the rate. Ask who eats the residual, and require the answer in the documents rather than in the press release.
Reported by CNBC on August 10, 2026.
