Nvidia pitches insurers on backing neocloud GPU loans as lenders balk

Nvidia has begun pitching insurers on a structure that would put them behind losses on loans to AI cloud companies buying its chips, according to Financial Times reporting that Bisnow carried Wednesday. Under the arrangement being discussed, insurers would compensate lenders when a neocloud defaults and the GPUs securing the loan cannot be resold for enough to repay the outstanding debt, though the talks are early and no agreements are in place, leaving the market with a proposal rather than a product.

For the early years of the AI boom, nearly all demand for its chips came from Amazon, Microsoft, Google and Meta, investment-grade buyers that funded purchases from their own balance sheets. Over the past 24 months a growing share of the growth has come from neoclouds instead, the GPU-specialist cloud providers that include CoreWeave, Core Scientific, Nscale and Lambda and that Bisnow describes as the fastest-growing segment of the data center market. Synergy Research Group has the group's revenue growing 223% year over year in the fourth quarter of 2025.

Neoclouds are also the tenants that have to borrow, nearly always financing both the GPUs and the data centers that house them—a tall order for young companies with short track records and a business model the coverage calls nascent and largely unproven. Lenders have balked at the billions involved, and the FT's account reads as Nvidia trying to make that debt digestible by moving part of the loss risk off a lender's balance sheet and onto a policy.

The structure addresses recovery, not payment. GPUs increasingly serve as the collateral in these financings, so what a lender recovers after a neocloud fails depends on what processors fetch on resale at the moment of failure rather than what they cost when the loan was made; insurance against insufficient resale proceeds is therefore a bet on a depreciation curve with very little history behind it, and fitting that curve is a judgment lenders have declined to make for themselves.

For anyone underwriting the sector, the credit issue is that a processor's residual value is set by how fast each new generation supersedes the last, a schedule detached from cap rates and replacement costs. Insurance written against a resale floor is a bet that the floor holds through a hardware cycle, and the risk a lender cannot price is that the collateral is worth materially less than the loan at the moment it matters, which is exactly the risk a policy of this kind would move to someone else.

A residual-value policy also concentrates that bet rather than spreading it, since the insured exposure is one borrower class and one collateral type and the secondhand GPU bid is a single market that every new generation resets. That suggests coverage of this kind would be written narrowly and priced accordingly, if it is written at all; the coverage reports no signed agreement and does not name an insurer at the table.

What the coverage does not describe is any slackening of demand for the chips themselves; it describes financing them, and the report says it has become harder for a growing share of Nvidia's customer base to reach the capital required to buy its products. Those same borrowers need to finance the shells that house the processors, and their short histories and unproven economics are what make lenders wary of computing capacity and the real estate around it in the same way, because a data center lease is worth what the tenant is worth and the tenant here is the same borrower whose chip loan started the conversation.

Capital is hunting data centers faster than creditworthy tenants appear. Data-center deals drove July's commercial real estate volume to a two-decade high, with a $33.8 billion month propping up an otherwise flat market, as PRED reported at the time. Blue Owl is in exclusive talks over Stack's Asia Pacific portfolio, where a $25 billion income bid sits against a build-out ask. Samsung committed $1 billion to KKR's Helix AI infrastructure platform, whose founding investors include Nvidia and Vistra. Mapletree's 22-building, 3.1 million-square-foot offering is asking the market whether buyers will pay for stabilized data center income or keep pricing the sector as a construction story.

The demand side carries its own warning. Bain & Co. estimates the industry must produce $6 trillion in annual revenue by 2031 to support the buildout, against as much as $1.8 trillion that existing consumer and enterprise AI can cover, a $4.2 trillion gap this publication covered in September. A chip vendor that sells today and finances its buyers' purchases is leaning on tenants whose own revenue has to arrive later, and the neocloud cohort is where that timing is tightest.

Nvidia is already on the equity side of that logic, as a founding investor in KKR's Helix platform, and the insurance talks would extend its involvement into the credit layer that funds its buyers. If an insurer can be persuaded to write the coverage, the FT's reporting suggests the effect would be to broaden capital access for some of Nvidia's riskier customers, on whose continued purchases Nvidia's own growth depends—a dynamic the coverage frames as a potential threat to the chipmaker. The talks are early, no agreement is in place, and no insurer is named. Until one signs, the resale value of a rack of processors remains a risk the lenders are carrying themselves.