Nvidia's CEO Said Old Chips Are Worthless. Now It's Guaranteeing $125 Billion They Aren't.

Jensen Huang said it plainly last November.

When Blackwell starts shipping in volume, you couldn't give Hoppers away. — Jensen Huang

This week, the company whose CEO said that signed agreements with six Wall Street firms — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR — to mobilize over $500 billion in third-party capital for AI infrastructure. Nvidia is providing residual-value guarantees of up to 25% on the GPU collateral backing those loans: a backstop of up to $125 billion on the premise that the chips hold their value. ^[1] The contradiction is not hidden. It is the structure. Nvidia now occupies four roles simultaneously, and each one tightens the conflict. First, it is the chip supplier. Nvidia designs and sells the GPUs that power the AI buildout, and demand has produced a 62% revenue increase and $31.9 billion in net income. ^[2] Second, it is the collateral guarantor. The $500 billion Wall Street partnership depends on lenders accepting GPUs as security. Nvidia's residual-value guarantee — covering up to 25% of the collateral's value — is what makes those lenders comfortable. Without it, a bank lending against a chip whose successor could render it obsolete would be taking a risk few credit committees would approve. With it, pension funds, insurers, and banks can treat AI compute as an infrastructure asset class. ^[1] Third, it is the financier. In July, Nvidia launched a revenue-sharing program called DSX that allows AI cloud providers and startups to acquire computing infrastructure and token credits without upfront capital, sharing revenue with Nvidia in return. Nvidia's CFO, Colette Kress, said the program bypasses "site selection, power procurement, construction and hardware bring-up." ^[3]

For model builders, inference providers, agentect platforms and enterprises scaling AI, it can mean faster access to full-stack accelerated computing without waiting through site selection, power procurement, construction and hardware bring-up — Colette Kress

Nvidia is now financing its own customers' infrastructure buildout through revenue-sharing rather than upfront capital, effectively underwriting their ability to acquire its chips. ^[3] Fourth, it is a venture investor in those same customers. Nvidia has committed roughly $20 billion to OpenAI and $10 billion to Anthropic — two of the largest buyers of its GPUs. ^[4] The arrangement has the shape of a loop: Nvidia sells chips to these companies, then invests in them, which can raise their valuations and help them raise more capital, which they can use to buy more Nvidia chips. When the OpenAI partnership was announced in October 2025, GMO's Tom Hancock compared it to vendor financing from the dot-com era.

I have to say the words ‘vendor financing’ do not carry nice reflections to somebody of my age. — Steve Austin

Separately, Nvidia has also struck a $20 billion licensing deal with Groq, a rival chipmaker — not a GPU customer but a competitor whose technology OpenAI had been courting as an alternative to Nvidia's own. ^[4] The entanglement extends beyond the customer base into the competitive landscape itself. The crux is the depreciation question. Nvidia's product-cycle decisions are the variable it controls that most directly determines whether hundreds of billions of dollars in collateral retain their value. When Huang said you could not give Hoppers away once Blackwell shipped, he was describing exactly the mechanism his company now guarantees against. The entity that decides when a chip generation becomes obsolete is the same entity that has promised Wall Street it will not become obsolete too fast. There is honest counter-evidence that GPUs have held value longer than the bear case assumes. CoreWeave, the GPU cloud provider, has signed contracts for Nvidia's A100 chips — launched in 2020 — extending into 2029, with pricing the company says is at or above where it was years ago. ^[5]

We recently signed an A100 contract that extends into 2029 at an attractive price. — Nitin Agrawal

Part of this persistence is physical: many existing data centers use air cooling and cannot support the liquid cooling that newer chips require, which creates structural demand for older-generation hardware that has nothing to do with technological competitiveness. ^[5] U.S. data center REITs have delivered 13.75% annualized returns over the past decade, the highest of any REIT specialization, and were outperforming before the AI boom began. ^[6] This evidence tempers the bear case but does not resolve the contradiction. It defers it. A physical constraint — air-cooled facilities that cannot be retrofitted — is not a permanent moat. It lasts only as long as the installed base of legacy data centers remains economically relevant. And the entity that controls the pace at which that relevance erodes is the same entity that has guaranteed it will not erode too quickly. The entity that guarantees the collateral's value is the same entity that decides when to make it worthless. No amount of Wall Street engineering changes who holds that switch.