Nvidia’s plan to mobilize more than 500 billion dollars of third-party capital for AI infrastructure is running into resistance on Wall Street, with lenders questioning whether the company’s GPUs can reliably serve as loan collateral over the multi-year life of a data center deal. Banking sources and credit managers told Reuters that lenders want stronger guarantees than Nvidia originally offered, and that some structures now under discussion would require the company to guarantee every deal or rely on revenue from investment-grade customers to cover the debt.
The plan, announced in August with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, set a target of more than $500 billion in financing for data centers, chip factories and power plants. The core idea was to build what Nvidia calls compute financing platforms, vehicles that would extend loans and leases to AI developers and cloud providers with Nvidia’s chips pledged as security. Nvidia does not write the checks itself, and the $500 billion is a cumulative, multi-year ambition, not money committed to any single fund or project.
Jensen Huang, Nvidia’s chief executive, has framed the thesis in one sentence: computing power is income. The financing plan is the practical expression of that idea, and the goal, in Nvidia’s telling, is to bring together the largest providers of long-term capital to fund compute the way they fund real estate or energy assets. Nvidia argues that top-tier GPUs can keep generating revenue for up to a decade, which is the period lenders would need to recover their principal on 10-year or longer data center projects.
The residual-value gap
The sticking point is the counterfactual. Banks traditionally underwrite AI hardware on a three-to-four-year depreciation schedule, pricing the asset on its useful life in a market where new chip generations arrive yearly. Nvidia’s decade claim relies on the argument that demand for AI compute will stay intense enough to keep older chips profitable well past the point a traditional lender would write them off. That gap between a chip’s accounting life and its claimed economic life is what the residual-value guarantee is meant to cover.
| Proposal element | Nvidia position | Lender position |
|---|---|---|
| Collateral basis | GPUs as long-term assets | Doubt about decade-long chip value |
| Residual-value guarantee | Up to 25 percent per deal | Too low, want stronger or universal coverage |
| Revenue duration | Chips earning for about a decade | Three to four year depreciation comfort zone |
| Credit assessment | Stays with capital providers | Agreed, but want more collateral certainty |
| Deals closed so far | Target announced, none signed | Structures under renegotiation |
Lenders are not yet ready to treat a data center full of GPUs the way they treat an aircraft fleet, where decades of resale history and standardized maintenance records make residual values predictable. Three banking sources who reviewed the plan but are not part of the original financing group told Reuters Nvidia may need to extend guarantees across all deals, or back each one with a revenue stream from an investment-grade technology customer, to make the structures bankable. Aircraft finance took most of a century to build that track record.
Analogies, precedents and nearby deals
Wall Street has closed adjacent deals that set the tone. CoreWeave, the AI cloud company built heavily on Nvidia hardware, pulled an $8.5 billion loan with GPUs supporting the facility. Anthropic has struck financing arrangements with Macquarie Asset Management and Singapore’s GIC to fund its compute purchases, though the value of those deals was not disclosed, and the company has warned that far more capital will be needed. BlackRock separately financed a Meta data center in Texas. The Nvidia platform is meant to generalize these structures, but each of those deals leaned on a specific customer’s contract, not on the chips themselves as an asset class.
Nvidia’s counter, in the words of Huang, is that its own guarantee is a minority position and the real underwriting still belongs to the lenders. The company plans to support individual projects with residual-value guarantees capped at 25 percent of the chips’ residual value, a share he has described as substantially smaller than in other compute financing arrangements. That framing puts the burden of the credit decision on the institutions putting up capital, but it does not resolve their core worry: without a reliable way to value a GPU in year six, they cannot size the loan against the collateral. Accusations that Nvidia is financing its own customers in a circular loop prompted part of this pushback, and the residual-value cap is partly Nvidia’s answer to that criticism.
Dan Rohinton, a portfolio manager at iA Global Asset Management, summed up the market reaction by lining up industrial names in Nvidia’s neighborhood on his radar, rather than treating the financing plan as settled. Tony Trzcinka, a senior portfolio manager at Impax Asset Management, was more blunt, telling Reuters that Wall Street remains far more conservative than the chipmaker’s own estimate of a decade of chip revenue.
The skepticism lands at a delicate moment for Nvidia’s stock, which has carried the AI trade for two years and now draws questions not about demand but about the financial engineering built on top of it. Reuters reporters noted lenders are comparing notes across multiple proposed deals, which suggests the terms seen so far are not isolated requests but a collective negotiating position. If the institutions insist on full guarantees or investment-grade revenue backing, the cost of capital for AI data centers rises across the board, and some marginal projects will not clear the bar.
Nvidia has not closed any transactions under the new financing platform as of early October 2026, so the plan remains a target, not a deal book. The company says no single customer commitment is embedded in the $500 billion and no timeline has been set. Whether lenders and Nvidia reach a workable middle ground could shape how broadly AI infrastructure gets financed in the year ahead, and at what cost of capital. The next test will be the terms of the first transaction actually signed, since that number, more than any statement from Huang, will establish what a used GPU is worth to a lender.