What it is
Residual-value risk is a standard concept in equipment and vehicle leasing, not a new invention. When a bank or leasing company finances an asset — a car, an aircraft, a fleet of servers — it often assumes the asset will still be worth a certain amount at the end of the financing term, and structures payments so the lessee covers only the difference between the asset’s starting value and that assumed future “residual” value. Residual-value risk is the possibility that the asset is worth less than assumed when the term ends: it depreciates faster, becomes technologically obsolete, or hits a soft resale market. Whoever is on the hook for that shortfall — the lessor, an insurer, or in some structures the original manufacturer — is said to be carrying the residual-value risk.
Accounting and banking regulators treat this as a distinct, quantifiable category of exposure. Under the IFRS 16 lease accounting standard, a “guaranteed residual value” is one where an unrelated third party promises the lessor a minimum recovery; an “unguaranteed residual value” is the portion the lessor simply has to hope for. In U.S. bank regulation, the Office of the Comptroller of the Currency has historically capped how much a national bank may rely on projected residual value — rather than the borrower’s own creditworthiness — when structuring a lease, precisely because residual-value assumptions have a history of being wrong in ways that surface only years later, when the asset comes back and the market has moved.
The risk is asymmetric in a specific way: it doesn’t show up on day one. A lease or loan can look fully collateralized at inception and still generate large losses at maturity, because the assumption embedded in the deal — what the collateral will be worth later — was never tested by an actual sale until the term expired.
Why it matters for AI governance and narratives
The concept has moved from a niche accounting and bank-supervision term to a load-bearing element of the AI infrastructure narrative because of how AI compute is now being financed. Nvidia’s roughly $500 billion financing framework, assembled with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, treats GPUs as collateral for infrastructure lending at a scale that requires banks and asset managers — historically wary of chip depreciation — to underwrite long-dated exposure to hardware whose useful life is contested. Nvidia has responded by offering to guarantee up to 25 percent of the residual-value gap on individual projects if resale or reuse value falls short. That is a direct, structural response to residual-value risk, and its existence is itself a signal: it tells the market that without some backstop, financiers would not treat GPUs as bankable collateral on the terms Nvidia wants.
This is where the concept becomes a narrative fault line rather than a technical footnote. Whether residual-value risk in AI compute is well-managed or systematically underpriced is contested along predictable ecosystem lines — skeptics like investor Michael Burry argue that GPU depreciation is being understated by tens of billions of dollars against Nvidia’s two-to-three-year product cycle, while Nvidia’s own framing points to sustained rental rates for older chips as evidence that value holds up longer than critics assume. Coverage that reaches for historical analogies — nineteenth-century railway finance, telecom fiber overbuild — is making an implicit argument about whether today’s compute buildout is durable infrastructure or a bubble whose collateral will not survive the next hardware generation. The term itself is now doing editorial work: which ecosystem invokes it, and in what tone, is a proxy for whether they believe the AI infrastructure boom rests on real, durable value.
Key facts and dates
Nvidia’s financing framework and its 25-percent residual-value backstop were announced in August 2026, with CEO Jensen Huang describing the guarantee level as “significantly lower” than typical compute-financing arrangements and citing rising H100 and B200 rental rates as evidence chips retain value longer than critics assume. The OCC’s lease-financing guidance for national banks has long applied a comparable 25-percent ceiling on how much residual value a bank may rely on in structuring a lease — a coincidence of numbers, not a direct comparison, since the two apply to different instruments and different eras of asset class. IFRS 16, the global lease accounting standard, formalizes the guaranteed/unguaranteed residual-value distinction that underpins how any such guarantee gets booked and disclosed.
Where to learn more
- Nvidia guarantees its own chips’ value to unlock $500 billion in AI infrastructure financing — The Decoder
- Comptroller’s Handbook: Lease Financing — Office of the Comptroller of the Currency
- IFRS 16 Leases — IFRS Foundation
- Why Jensen Huang’s $500 billion AI financing plan faces a big risk from China — CNBC