Nvidia has signalled to the market that it is willing to absorb part of the loss if its own GPUs end up worth less than promised. This isn’t marketing language. It is a contractual guarantee, capped at 25 percent of any given transaction and assessed on a project-by-project basis.
That mechanism sits underneath the headline figure: letters of intent signed with six financial firms to mobilise more than $500 billion in third-party capital for data centers, chip factories and power plants. Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR are the partners involved.
The market was unimpressed. Nvidia shares slipped roughly 1.4 percent once the deal came to light, wiping out more than $70 billion in market cap.
What the guarantee actually covers
Should the resale or reuse value of installed hardware land below expectations when a financing term ends, Nvidia picks up part of the shortfall — as much as 25 percent of the transaction. In effect, the chipmaker is shouldering a portion of the depreciation risk attached to its own products.
Nvidia CEO Jensen Huang described that share as “significantly lower” than what appears in comparable compute financing arrangements. The credit assessment itself — judging the customer, demand, utilization, cash flow and residual value — remains the responsibility of the capital providers.
The distinction matters. Nvidia takes the tail risk on the asset. The borrower is underwritten by someone else.
The number is a target, not revenue
Writing on X, Huang framed the move as a pivot away from one-off projects toward repeatable financing platforms, with “AI factories” financed the way power grids or transportation networks are. Many AI companies, he noted, have all the demand for compute they could want and no route to raising capital at the scale required.
He was blunt about what the $500 billion is not. The figure is an aggregate target stretched across years — not Nvidia revenue, not a single fund, and not a commitment tied to any one customer.
Terms, individual commitments and a timeline all went unpublished. That leaves a great deal of unspecified structure propping up a very precise number.
The circularity question Huang couldn’t dodge
Nvidia routinely helps its partners take on debt, and that debt comes back around as Nvidia revenue. Huang tackled the accusation head-on, and the residual-value guarantee is his rebuttal: genuine risk moved back onto Nvidia’s own balance sheet, rather than vendor financing repackaged as demand.
A similar guarantee is currently under negotiation for a 10-gigawatt data center in Ohio leased to OpenAI.
Michael Burry called this the fraud, and Huang is arguing the other side
Investor Michael Burry has labelled the hyperscalers’ depreciation practices “one of the more common frauds of the modern era.” His case: GPUs become obsolete far too quickly to support five-to-seven-year useful lives, given that Nvidia itself releases a new generation every two to three years. He pegged the understatement at around $176 billion for the 2026 to 2028 period alone.
Huang argues the reverse. He points to the A100, launched in 2020 and still in commercial service six years on, with an economic lifespan he sees running toward a decade — helped along by CUDA, which he says keeps improving hardware already in the field.
Rental pricing is the evidence he leans on. Annual H100 contracts climbed from $1.70 per GPU-hour in October 2025 to $2.35 by March 2026, while B200 capacity trades between $5.30 and $7.05.
Rents rising on a five-year-old part is a real data point. It is also precisely the sort of price signal that persists right up until supply catches up.
The numbers everyone else is working with
Morgan Stanley forecasts $3.5 trillion in hyperscaler spending between 2026 and 2028. Apollo president Jim Zelter pegs the overall investment requirement above $8 trillion. Set against $8 trillion, a $500 billion financing target is a fraction of the picture rather than the picture itself.
In its July Financial Stability Report, the Bank of England cautioned that the pace here is historically unprecedented, and that a shock landing on highly leveraged AI companies could spread through global financing conditions and set off a credit crunch. Banks and private credit firms, the report added, can see little of their indirect exposure.
That final point is the one to hold onto. The residual-value guarantee puts a lender’s downside on paper in legible terms — 25 percent, project by project. What no one has disclosed is who is carrying the remaining 75 percent, or how much of it loops back through the same six firms.