Nvidia Wants Wall Street to Lend $500 Billion Against Its Chips. Nobody Can Agree How Fast Those Chips Hit Zero.
In August, Nvidia said it was lining up as much as $500 billion in Wall Street financing for AI data centers — secured not by buildings or contracts, but by the GPUs themselves. The whole plan rests on one number nobody can agree on: how many years before an AI chip is worth nothing.
Drag the "how long does a GPU last?" slider
A simplified model: $500B of financing amortizes in a straight line over a hypothetical 5-year loan term (dashed blue line). GPU collateral value declines in a straight line to zero over whatever "useful life" you set (solid red line). Wherever collateral value falls below the loan balance still owed, that's an uncovered gap — shaded red. The dashed grey line marks Nvidia's publicly stated backstop ceiling, scaled to match its real ~25%/$125B figure. This is an illustrative model, not the real deal's undisclosed terms.
The plain version
Imagine financing a fleet of delivery trucks with a bank loan where the trucks themselves are the collateral. If you stop paying, the bank repossesses the trucks and sells them to cover what's owed. That works fine as long as everyone roughly agrees on how fast a truck loses value. Now swap "trucks" for "AI chips," and you've got Nvidia's new financing plan — except nobody agrees on the depreciation schedule at all.
In August 2026, Nvidia announced it was working with six of Wall Street's biggest firms — Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs, and KKR — to arrange up to $500 billion in financing for AI data centers. The twist: much of the collateral backing that debt isn't a building, it's racks of GPUs. CEO Jensen Huang said Nvidia itself might guarantee up to a quarter of any shortfall, around $125 billion, if the chips end up worth less than expected.
Here's the catch. Cloud giant Meta currently assumes its servers keep earning their value for 5.5 years. Amazon just shortened its own assumption to 5 years, because AI hardware is improving so fast that older chips fall behind. Investor Michael Burry has gone much further, arguing publicly (though he hasn't opened his books to prove it) that a real AI chip is economically dead in 2 to 3 years, once a faster, cheaper model replaces it — which would mean years of industry profits have quietly been overstated.
There's already a live test case: CoreWeave, an AI cloud company, closed an $8.5 billion loan this year backed by GPUs plus a Meta contract — the first such deal ever rated investment-grade. If it holds up, it's the template for the $500 billion plan. If GPU values crack faster than the loan gets paid down, somebody ends up holding chips worth less than what's still owed on them.
The expert version
The underlying risk is a maturity/value mismatch between an amortizing liability and a depreciating asset — structurally identical to the risk sitting under any asset-backed security, from mortgage bonds to aircraft leases, except the asset here is semiconductor hardware on an unusually compressed replacement cadence (Hopper → Blackwell → Rubin, roughly annual).
In August 2026, Nvidia disclosed non-binding arrangements with six asset managers (Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs, KKR) to mobilize up to $500 billion in financing capacity for AI infrastructure. Structurally, special-purpose vehicles raise debt against compute capacity — GPUs plus associated customer contracts — then lease that capacity to operators. Huang stated Nvidia could backstop "up to 25% of an opportunity," implying a ceiling near $125 billion, though Nvidia has publicly denied any binding $125B commitment, and reporting is split on whether $500B represents committed capital or an aspirational capacity target.
The precedent instrument is CoreWeave's $8.5 billion delayed-draw term loan, closed March 2026: the first GPU-backed facility to receive investment-grade ratings (Moody's A3, DBRS A-low), structured through a bankruptcy-remote special-purpose subsidiary, collateralized by GPU servers plus a multibillion-dollar Meta compute contract.
The open variable is the useful-life assumption driving straight-line depreciation, which determines how fast book collateral value falls relative to loan amortization. Public hyperscalers disagree with each other: Meta extended server/network useful life to 5.5 years effective January 2025 (cutting 2025 depreciation expense by an estimated $2.9 billion), while Amazon shortened a subset of servers from 6 to 5 years, citing accelerating AI hardware turnover, taking a one-time accelerated-depreciation charge reported at roughly $700–920 million across outlets (figures vary by source). Michael Burry has argued publicly — a claim CNBC said it could not independently verify — that true AI-chip economic life is closer to 2–3 years, which would imply hyperscalers understate cumulative depreciation by tens of billions annually. Countervailing evidence: Bernstein analyst Stacy Rasgon notes 5-year-old A100s still carry solid margins, and CoreWeave renewed an expiring H100 contract within 5% of its original price — real-world resale/lease value decaying slower than the bear case assumes. Which curve is closer to true determines whether $500 billion of GPU-backed debt behaves like investment-grade paper or like 2007-vintage subprime collateral.
Why it matters for tech + supply chain: If AI's physical collateral ages faster than its paperwork assumes, the first crack in a downturn won't show up in a chip fab — it'll show up in a loan covenant.
Why it matters for tech + supply chain: Hundreds of billions in planned fab capacity, power contracts, and chip orders are now downstream of an accounting estimate — useful life — that two hyperscalers can't even agree on with each other.