Nvidia's Chip-Backed AI Financing Meets Wall Street Skepticism

Nvidia wants the world to treat AI chips as an asset class you can borrow against. Wall Street is not convinced yet, and the gap between the two views is about to show up in what it costs to rent compute.

What happened

According to Reuters, lenders and credit managers are demanding stronger guarantees than Nvidia originally offered under the $500 billion financing programme it announced in August with Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR.

The disagreement is about time. Nvidia CEO Jensen Huang has argued that top-tier GPUs keep generating revenue for up to a decade, and the company points to third-party valuations suggesting its GB300 NVL72 systems carry a nine-to-ten year useful life. Lenders are pricing the hardware on a much shorter clock.

"Banks typically underwrite GPUs over a 3-4 year depreciation schedule," Impax Asset Management portfolio manager Tony Trzcinka told Reuters. S&P Global Ratings director Andrew Chang said GPUs have so far proved to work well beyond five years, but that the agency still takes "a conservative view of the value of those chips."

Nvidia's own backstop is limited. The company has said some deals could carry no more than a 25% residual value guarantee. A spokesperson told Reuters that "AI compute is a productive, durable and fungible asset that can support long-term financing."

Why it matters

Every projection about AI infrastructure rests on an assumption about how fast the hardware underneath it loses value. If lenders underwrite GPUs over three to four years instead of ten, that discount lands in lease rates, then in the price of inference, then in the margin of every product built on top of a model.

The deals already closed show what lenders actually trust. CoreWeave's $8.5 billion GPU-backed facility earned an A3 rating mainly because it leans on Meta's contractual payments rather than the chips themselves. Broadcom backstopped more than 80% of a $35 billion structure financing Anthropic's compute. In both cases the credit is really underwriting a customer, not silicon.

The builder angle

Three practical takeaways. First, long-term capacity contracts are becoming collateral, so teams that can commit to multi-year spend will get better compute economics than teams renting month to month. Second, audit your provider's depreciation assumptions: a two-year difference in assumed GPU life moves a multi-year inference cost model more than most price cuts do. Third, financing capability is turning into a vendor advantage, because the companies able to underwrite your compute are increasingly the ones selling it. It is the same pattern we tracked in Nvidia's push to move agent guardrails into the runtime, where control over the layer underneath becomes the real product.

Nvidia's hardware may well hold value for a decade. It just is not an asset class yet, and the difference is measured in basis points.