Nvidia Will Cover Part of the Loss If Its Own Chips Lose Value. That Is How You Raise 500 Billion Dollars.
Six of the biggest finance houses want to fund AI data centres. To make the numbers work, Nvidia is guaranteeing up to a quarter of what its hardware will be worth at the end.
Nvidia has signed letters of intent with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise more than 500 billion dollars of outside money for data centres, chip plants and power stations. The idea, as chief executive Jensen Huang described it, is to stop treating each AI data centre as a one off deal and start financing them the way you finance a power grid or a railway. Investors were not immediately charmed: Nvidia’s share price fell about 1.4 percent after the Financial Times broke the story, wiping out more than 70 billion dollars in market value.
The interesting clause is buried in the structure. To reassure lenders, Nvidia will offer residual value guarantees on individual projects: if the installed hardware is worth less than expected when the financing term ends, Nvidia covers part of the shortfall, up to 25 percent of a given transaction, assessed project by project. In plain terms, the company selling the chips is underwriting a slice of the risk that its own chips go stale. Huang says that share is well below what is normal in comparable compute financing, and that the actual credit judgement stays with the banks.
Why that clause exists at all is the real story. Investor Michael Burry has argued that the big cloud companies are writing off their graphics cards over five to seven years while Nvidia replaces them every two to three, and that depreciation is therefore understated by roughly 176 billion dollars between 2026 and 2028. Huang’s answer is a direct rebuttal: the A100 launched in 2020 is still earning money six years later, CUDA keeps making old hardware more useful, and rental prices are rising rather than collapsing. He cites H100 annual contracts moving from 1.70 dollars per GPU hour in October 2025 to 2.35 in March 2026, with B200 capacity at 5.30 to 7.05.
Both men are describing the same uncertainty from opposite ends, and the guarantee is what happens when you have to price it. Morgan Stanley expects 3.5 trillion dollars of cloud company spending between 2026 and 2028. Apollo’s president puts the total need above 8 trillion. The Bank of England warned in its July Financial Stability Report that the pace is historically unprecedented, that a shock to highly leveraged AI firms could tighten global credit, and that banks have limited visibility into their own indirect exposure.
What this means for you: nothing about your chatbot changes today. But this is the machinery that decides whether AI stays cheap. Free tiers, flat rate subscriptions and 20 dollar plans exist because someone else is financing the hardware, and structures like this one are how that financing gets extended. If the residual value bet goes well, prices stay low longer. If it does not, the correction shows up as higher prices and thinner free plans, and, per the Bank of England, possibly somewhere more uncomfortable than your subscription. Worth watching, not worth panicking about.
Sources
Small Models Read Old Books Better Than Big Ones, and It Costs Two Dollars per Thousand Pages
Hugging Face and EleutherAI tested 14 open text recognition models on historical pages. The winner has 3 billion parameters and beats a model three times its size.
