Nvidia wants investors to finance AI compute like infrastructure. Whether that holds up depends on how long the machines keep earning, and who takes the loss when they age faster than the loans.
Nvidia wants Wall Street to treat AI compute as a new asset class. On Monday, alongside Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, it announced financing platforms meant to raise more than $500 billion of outside capital for AI “factories.” Those are the data centers that train and run AI models. CEO Jensen Huang argues that this compute can be an “investable asset class,” because it keeps earning and can be reused across many customers. The harder question is who takes the loss if the hardware ages faster than the loans are paid off. This isn’t mainly about whether AI demand is real. It’s about whether fast-aging machines can be financed like durable infrastructure.
What Nvidia Actually Announced
Let’s start with what it isn’t. These are memorandums of understanding (MOUs), not signed contracts, and Nvidia says each project still needs a final agreement. There’s no timeline yet, no word on how the money splits among the six firms, and no first project named. The $500 billion is a target for capital to be raised over time. It isn’t Nvidia’s revenue, and it isn’t one fund or one customer. The likely borrowers are AI labs, big enterprises, and the cloud companies that rent out computing.
The key word is “independent.” Nvidia says the six firms will judge each deal on their own. They weigh the customer’s demand, how hard the hardware runs, the cash it makes, and what it’s worth secondhand. Nvidia supplies the computing platform; the investors decide what to fund. That lets Nvidia expand its customers’ buying power without putting every project on its own balance sheet.
The Real Test Is Who Takes The Loss
An AI data-center loan rests on two things: who has agreed to pay for the computing, and what the equipment is worth if that customer walks away. The second number is residual value. It tells a lender how much the hardware itself protects them once the revenue stops.
In his essay, Huang says Nvidia may offer “residual-value support for up to 25% of an opportunity,” decided case by case. He calls it limited, and says it adds to independent underwriting rather than replacing it. He hasn’t said how it works, or who is on the hook first if a deal fails. So until the contracts are public, this isn’t a guarantee. We shouldn’t assume Nvidia takes the first loss. Still, it tells you something. Outside money may only show up because Nvidia is willing to cover part of the downside.
I’ve sat through AI-infrastructure financing reviews, and demand is only the first question. The harder one is who owns the boxes when the lease ends, and what a three-year-old chip fetches then. Nvidia offering its own credit makes that money easier to raise. It also means “independent underwriting” should be judged by the contracts, not the press release.
Rental Prices Don’t Prove Residual Value
Huang’s evidence for durability is rental pricing, and it needs a careful read. He notes that a one-year rental for the H100 chip rose from about $1.70 an hour in October 2025 to $2.35 in March 2026. That shows what the chip earns now. It doesn’t show what it would sell for. The period matters too. By Silicon Data’s tracking, the median price to rent an H100 from a big cloud ran near $9.34 an hour in the second half of 2024 and about $6.26 a year later. Marketplace and smaller-cloud rates sit elsewhere again. There’s no single H100 price.
Rental income and resale value answer different questions. Nvidia’s own case is that its platform can move from one customer to another, and that CUDA software keeps improving output on chips already installed. That can support the cash flow, but it doesn’t prove a strong resale price. Airlines borrow against planes the same way, because the contracted revenue pays down the loan before resale value matters. AI lenders need that same fit between the loan and the hardware’s working life. The risk is a new chip generation landing before the debt is paid off.
Amazon’s Accounting Sets A Shorter Clock
Amazon’s books offer a warning. Effective January 1, 2025, it shortened the estimated useful life of some of its servers and networking gear from six years to five. Its reason, in its own words: “the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.” Amazon says the change added about $1.4 billion to its 2025 depreciation and cut net income by roughly $1 billion, mostly at AWS.
Read it carefully, though. Amazon didn’t write down Nvidia chips, and its filing doesn’t name them. What it shows is that one of the biggest AI operators decided some of this gear wears out faster than it used to assume. Investor Michael Burry went further in November 2025, estimating that big cloud firms were understating AI depreciation by about $176 billion from 2026 through 2028. That’s his estimate, not a reported loss. Nvidia argues the opposite. It says A100 chips from 2020 still draw multi-year commitments that can stretch their useful life toward a decade. That gap is exactly what lenders have to price.
The Contracts Create The Asset Class
Calling AI compute an asset class doesn’t make it one. The contracts do. The debt already flowing into AI shows why executives need better terms, not better labels. Before signing off on GPU-backed financing, ask four things: who has committed to use the capacity, whether the loan clears before the next hardware refresh, who can redeploy the machines, and who takes the first loss if resale value comes up short. The same question drives both the overbuilding worries at Meta and the fight over power and transformers: can the owner prove the hardware will earn before it ages? Demand can be real and the collateral still disappoint. Nvidia’s $500 billion push will hold up only if the cash flow outruns the aging. That’s the test as the first deals move from handshake to signature.
