Bargaining chips: Nvidia is the bank of AI

Nvidia took three decades to reach a valuation of $1 trillion, then doubled it in nine months and crossed $5 trillion less than two years later. But the chip designer now faces a question that would have seemed absurd during the early AI boom: if demand for its processors is so insatiable, why is the company effectively becoming a bank for its own customers?

The financial engineering behind the chip sales

The core problem is that compute—the processors and memory chips that make up an AI data center—is staggeringly expensive. A single one-gigawatt data center costs nearly $50 billion. The hyperscalers like Amazon, Google, and Microsoft can afford that spending, but smaller AI labs and startups bleeding cash cannot. So Nvidia has begun deploying its balance sheet to help them pay for its own products.

In mid-August, Nvidia agreed to guarantee up to $105 billion so that OpenAI could lease a vast data center in Ohio. A week earlier, it announced a partnership with six Wall Street investment firms including Goldman Sachs, Blackstone, and BlackRock to raise over $500 billion for an AI infrastructure fund built on Nvidia's equipment—with Nvidia potentially backstopping around 25% of those deals by guaranteeing the value of its chips. When you add up all the backstop commitments, purchase commitments, and equity investments in startups, the total exceeds $350 billion.