Nvidia Seeks Insurers to Back Loans Secured by Its Own Chips

A loan can be secured by chips. That is no longer new.

The new question is this: if the loan goes bad and the chips cannot be sold at the expected price, who pays the difference?

According to a Financial Times report published on September 29, Nvidia is looking to insurers for the answer.

What the Talks Cover

The FT reports that Nvidia is working with reinsurance broker Howden Re on a possible structure. The target is loans to smaller cloud providers known as "neoclouds."

These companies buy Nvidia chips, pledge them as collateral and rent out the computing power. If a borrower defaults and the seized chips fail to cover the debt, the insurance would make the lender whole.

Nvidia has shared chip depreciation rates and expected future compute prices with at least one insurer. It has also considered a second structure in which insurers pass the risk on to hedge funds.

Nvidia said the talks are at an early stage and may not lead to deals. The core line of its statement:

"AI infrastructure is an investable asset class because it's uniquely productive, durable and fungible."

Nvidia Already Carries Part of the Risk

The CFO commentary Nvidia filed with the SEC on August 26 shows how large its guarantees have become. Maximum exposure on guarantees to AI cloud partners is $3.5B.

The guarantee for SB Energy and OpenAI data centers is capped at $105B. It takes effect in phases, with the first expected in fiscal 2029. Total maximum exposure is $108.5B.

On top of that sit $36B in AI cloud revenue-share agreements and $29B in cloud service commitments.

A company that sells chips had already started vouching for its customers' debts. The insurance talks are the next link in that chain.

This looks like an insurance story.

It is really a story about pricing collateral.

The Risk Sits in the Chip's Resale Price

In a chip-backed loan, a bank looks at two things. One is the borrower's ability to pay. The other is what the chips will fetch if the borrower fails. The second is uncertain for AI chips, because resale value depends on rental income.

Market data show H100 hourly rents falling from $7-10 in 2023 to $2-4 by late 2025. Compute Exchange's April data put a new H100 at $25,000-40,000 and a used one at $15,000-28,000.

Insurance does not remove the risk. It prices the risk and moves it to another balance sheet.

For Nvidia, the result is simple. Smaller buyers get cheaper debt, and Nvidia sells more chips without enlarging its own guarantees.

The Part That Does Not Fit

The company supplying the depreciation curve is the company selling the chips. The policy's price rests on the seller's own estimate of how well its product holds value.

According to FT insurance correspondent Lee Harris, traditional insurers are already near their maximum exposure limits for AI companies. That helps explain why the risk would be passed on to hedge funds.

The depreciation debate is not settled either. In 2025 Amazon cut server useful lives from six years to five; Meta lengthened them in the same period.

The Stock Did Not Price the News

Nvidia traded 0.9% higher before the open on September 29 and, according to reports, closed down 0.3%. That day's agenda also included the rise in 30-year Treasury yields and the previous day's buyback announcement.

The chart is not evidence of direction; it lines up three months of financing announcements against the price. Over the same period, the Nasdaq 100:

The lack of movement is itself a finding. This is still a conversation, not a policy, with no size or price attached. Nvidia's latest quarterly figures are in our earnings analysis.

This article is based on the Financial Times report of September 29, 2026 and its coverage by The Next Web, Yahoo Finance and Benzinga. It also draws on Nvidia's CFO commentary filed with the SEC on August 26 (Form 8-K) and on used-GPU pricing data from Hashrate Index and Compute Exchange. The talks rest on sources who spoke to the FT; Nvidia said no deal has been reached. Rental and resale prices are market compilations, not official data.