Nvidia Wants Wall Street to Treat Chips Like Aircraft. Lenders Are Pricing Them Like Used Cars.

The version of this story that got the headlines is simple: Wall Street is skeptical about Nvidia's $500 billion financing plan. That framing is true, and almost useless. The fight is not about whether the plan works. It is about a single number that nobody outside a credit committee has been able to see clearly, and that number now sits underneath a meaningful share of the AI buildout.

The number is the useful life of a GPU.

The plan, and the part everyone skips

On August 10, Nvidia announced memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to create what it calls compute financing platforms, targeting more than $500 billion of third-party capital. The idea is to treat AI chips the way the market treats aircraft: expensive, long-lived equipment that can be leased and borrowed against. Jensen Huang framed it plainly. "In AI, compute is revenue," he wrote, and Nvidia compute is "broadly adopted, flexible across models and workloads, fungible and transferable across customers and operators."

The detail that matters sits below the headline. Nvidia said some deals could carry no more than a 25 percent residual value guarantee. That is a promise about what the hardware will be worth when a financing term ends, capped and assessed project by project. It is a deliberately modest commitment, and it is exactly what lenders have started pushing back on.

The gap between a decade and three years

Reuters reported on October 1 that some lenders want stronger guarantees than Nvidia outlined, and that deals already in the pipeline are likely to carry them. The reason is not philosophical. It is arithmetic.

"Banks typically underwrite GPUs over a 3-4 year depreciation schedule," said Tony Trzcinka, a senior portfolio manager at Impax Asset Management. "That is different than Nvidia which argues top-tier GPUs can earn revenue for a decade."

Andrew Chang, a director at S&P Global Ratings, put the same tension more carefully. "Nvidia would imply that the GPUs work well north of five years, and that actually has been proven to be true thus far," he said. "Yet we take a conservative view of the value of those chips." Loren Moran, a fixed income portfolio manager at Wellington Management, which runs $1.3 trillion, described the consequence in the language of price rather than principle: "you're going to be a lot pickier about the levels that you need to get compensated for to take incremental risk."

That is the whole story in one sentence. Change the assumed life of a chip from ten years to four and you have not changed the chip. You have changed the cost of capital for every company that wants to own one.

Nvidia's counterargument is not thin. It points to third-party studies showing major cloud companies extending server depreciation from three-to-four years to five-to-six, and to a Barkr finding that its GB300 NVL72 systems could have a useful life of nine to ten years. Huang has pointed to the A100, introduced in 2020, still in active commercial use. The trouble is that none of this rests on a resale market with enough history to underwrite against. A used jet has decades of transactions behind it. A used accelerator has a handful of generations and a supply chain that has spent three years unable to keep up with demand, which flatters every residual assumption in the model until the day it does not.

The precedents cut the other way

If you want to know what the credit market actually believes, stop reading forecasts and look at what has already been priced.

Brian Gelfand, co-head of global credit at TCW, which manages more than $200 billion, made the point directly: "The precedent transactions so far would suggest that the creditor community does not subscribe to long average lives for these assets."

He is right about the pattern. CoreWeave closed an $8.5 billion facility this year, described as the first investment-grade GPU-backed loan. It is rated A3 largely because lenders are relying on Meta's contractual payments, not on the chips. Broadcom backstopped more than 80 percent of a $35 billion structure supporting Anthropic's compute. Nvidia itself provided a residual value guarantee for SB Energy's Ohio data centre project. In each case the debt was made financeable by something other than the hardware: a creditworthy tenant, a vendor's balance sheet, or both.

Nvidia's own Ohio arrangement shows the template at full scale. The company's August 17 filing discloses residual value guarantees covering roughly 4.25 gigawatts of OpenAI leases at the PORTS-Pike campus in Pike County, Ohio, with an aggregate payment obligation capped at $105 billion. OpenAI is the tenant on a 20-year lease. Nvidia is putting $1.5 billion of real cash into SB Energy, the landlord, and holds an option on another 3.8 gigawatts. Payments cannot begin before ready-for-service conditions are met, expected from 2028. The guaranteed minimum values, the number that actually determines what Nvidia would owe, are left undisclosed.

Read the fine print and the instrument is less a guarantee of rent than a floor under resale value. The tenant has agreed to reimburse Nvidia for anything it pays, which is worth a great deal in the scenario where the tenant is solvent and considerably less in the one the guarantee was written for.

The tension is now system-wide

Two things landed alongside the Reuters story that belong next to it.

On October 1, the Bank for International Settlements published a bulletin on circular relationships among AI firms. Its numbers are blunt. Between 2021 and 2025, 55.2 percent of incoming investment into AI companies came from other AI companies, and 46.4 percent of AI-to-AI deal value also involved a commercial supply relationship between investor and target. The BIS notes that residual value guarantees are typically classified off balance sheet and matter most in exactly the conditions where the guarantor can least afford them. It also flags the 1990s telecom parallel, where equipment vendors financed the network operators buying their gear, booked the loans and the revenue, then absorbed both losses when demand did not arrive.

On October 2, Bloomberg reported that Broadcom's syndicate is assembling $60 billion of chip financing tied to Anthropic and others, including a $42 billion senior secured tranche. That deal is being watched as a live test of whether credit markets still want to fund AI hardware on these terms.

Both stories point the same direction. The question is no longer whether the AI buildout is real. It is who holds the depreciation when the equipment ages, and at what price that risk changes hands.

What this means if you build hardware

For anyone with a hardware roadmap, the read-through is that financing assumptions are turning into engineering constraints. If the cost of capital for your customers now depends on a residual value story, then the resale market, the lease structures and the refresh cycles around your product are no longer somebody else's problem. They are inputs to your own demand forecast.

That cuts both ways. A supplier willing to underwrite its own product's residual value can unlock demand that would otherwise sit unfunded. It also means the supplier is exposed to the resale price of the thing whose supply and cadence it controls. That is a new kind of balance sheet risk for companies that spent decades thinking of themselves as manufacturers.

The supply chain complexity behind financing a hardware product in the middle of an AI-driven capital boom is the kind of problem that does not show up in a spec sheet. At DMC, we work with hardware companies navigating exactly these constraints: sourcing strategy, cost modeling, and production ramp planning when the financing terms move as fast as the silicon. Need help stress-testing your hardware roadmap against this? let's talk.