Nvidia's $500 Billion Data Center Plan Is Really About Keeping Old Chips Valuable
Six of Wall Street's biggest names have pledged up to $500 billion to build AI data centers. The headline number is striking. The real story is how Nvidia is betting its own money that its chips hold their worth over time.
Key points
- Six major financial firms, including Goldman Sachs, BlackRock, Brookfield and KKR, pledged up to $500 billion to fund AI data center construction, with Nvidia's backing.
- Nvidia agreed to cover up to 25% of any gap if its GPUs, the specialised chips that power AI, lose value faster than expected when used as loan collateral.
- The arrangement creates what finance professionals call "wrong-way" risk: Nvidia's costs rise precisely when its revenues are likely falling.
- Nvidia CEO Jensen Huang argues AI computing hardware should be treated like long-term infrastructure, similar to railroads, not disposable electronics.
- The plan is designed to kickstart a second-hand market for used AI chips, potentially lowering costs for startups and smaller businesses.
Six of the biggest names in finance, Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, have said they'll commit up to $500 billion to build AI data centers. Nvidia is backing the deal. When we first reported the arrangement on 13 August, the focus was on where the money comes from. The part that matters more for ordinary businesses is quieter: Nvidia is trying to build a healthy market for used AI hardware.
What did Nvidia actually agree to?
Nvidia promised to partially guarantee the value of its own chips when those chips are used as collateral, meaning security against a loan, similar to how a house backs a mortgage. If a data center owner defaults and the lender can't sell the chips at book value, Nvidia covers up to 25% of the shortfall.
Nvidia CEO Jensen Huang, writing on X after bond markets reacted nervously, framed the move as bringing "independent, long-term institutional capital into the AI infrastructure market." The key distinction from simply lending customers money to buy your own product is that Nvidia shoulders only a slice of the risk while the big financial firms shoulder the rest.
That distinction matters because some commentators drew comparisons to Lucent Technologies, a telecoms equipment maker that collapsed after the dot-com bubble burst when it lent customers cash to buy its own gear. Nvidia insists the structure here is different. The bulk of capital and risk sits with the institutional investors, not on Nvidia's balance sheet.
What is the risk for Nvidia?
Financiers call it "wrong-way" risk. Nvidia's obligations grow at exactly the moment its revenues are likely shrinking. If demand for AI chips cools, the chips used as collateral lose value at the same time Nvidia's being asked to cover the gap.
Huang's counter-argument is that AI computing is long-lived infrastructure, not a gadget that goes obsolete in three years. He calls his servers "AI factories" and compares them to railroads or airlines: assets that get passed from operator to operator rather than thrown away.
"When needs change, the factory can be used by another customer or another cloud," Huang wrote on X.
What does this mean for startups and smaller businesses?
If Huang's vision holds, used Nvidia hardware becomes easier to buy at lower prices. A second-hand market for AI chips could mean a small company or a university research team accesses serious computing power without paying top dollar for brand-new equipment.
Think of it like a certified used-car market: the manufacturer's stamp of confidence keeps resale prices from collapsing, which also makes buyers more willing to purchase in the first place.
Survivorship bias is worth naming here. We hear Huang's optimistic version. The scenario where AI demand softens and chips flood the market, making Nvidia's guarantee costly, doesn't get a press release.
Here's the judgement on this one: the secondary GPU market is the most underreported part of this story. If Nvidia actually makes it work, the real winners won't be the six Wall Street firms, they'll be the mid-sized businesses that've been priced out of serious AI compute entirely. Watch used-GPU prices over the next twelve months. That number will tell you whether Huang's bet is paying off long before any earnings call does.
