Nvidia's Bank
In March, Jensen Huang said Nvidia was probably done writing big checks to AI labs. The money never stopped. It changed form, from equity stakes into loan guarantees and financing deals, and now roughly a quarter of Nvidia's business comes from customers it helps pay for.

The decision
On March 4, Jensen Huang told a Morgan Stanley conference that Nvidia’s investments in OpenAI and Anthropic would “probably” be its last big equity checks into AI labs. The headlines read it as a retreat.
It wasn’t. The money changed shape.
On the August 26 earnings call, CFO Colette Kress laid out the new version. Nvidia has put nearly $50 billion into AI labs. It has more than $500 billion in partnerships committed, still subject to final agreements. It has financing platforms set up with 6 of the largest investment firms in the world: Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR. Reports put its customer financing guarantees at about $108 billion.
Then she gave the number that matters. Labs Nvidia helps finance will make up roughly 25% of its business next year.
The biggest single example is in Ohio. In August, Nvidia agreed to back OpenAI’s data center campus in Pike County, which is planned to reach 8 gigawatts. Nvidia guarantees lease and power payments through SoftBank’s SB Energy and the value of the infrastructure if OpenAI defaults. The guarantee came in at $105 billion. Reports earlier in the summer had it at $250 billion, and Nvidia’s stock fell about 4.5% in a day on that number.
Nvidia stopped buying stakes in its customers. It started co-signing their leases. That is the decision.
The strategic why
On the surface, this looks like Nvidia funding its own sales, and it partly is. But look at who gets the money.
Nvidia’s biggest threat isn’t AMD. It’s its own largest customers. Google runs on its own TPUs. Amazon has Trainium. Microsoft and Meta are building chips too. Every dollar those giants spend on their own silicon is a dollar Nvidia never sees, and every chip they design gives them leverage when they negotiate with Nvidia.
The customers Nvidia finances look different. OpenAI. Neoclouds like CoreWeave. Data center developers like SB Energy, where Nvidia’s deal is for Nvidia-only hosting. These companies don’t build chips and aren’t likely to start. A dollar of Nvidia credit to them turns into Nvidia demand, and almost nothing else.
That’s the chess move. Huang can’t stop the giants from building their own chips. He can grow the part of the market that never will, and he can pay for that growth with guarantees instead of cash.
The form matters too. A $30 billion equity check shows up on Nvidia’s balance sheet, gets marked to market and invites questions about circular deals. A guarantee is a promise. It costs nothing until someone defaults. Partnering with Apollo and Blackstone means outside money funds the build while Nvidia supplies the confidence. Nvidia lends its credit rating, not its cash.
And there’s timing. OpenAI and Anthropic are both heading toward IPOs. Once they’re public with their own capital, Nvidia loses its seat at the table. Financing deals signed now lock in years of hardware purchases before that happens.
It’s also why the $150 billion buyback and the guarantees fit together. Nvidia makes so much cash that it can return a record amount to shareholders and still stand behind more than $100 billion of its customers’ obligations. No other chip company can make that offer. A rival can match Nvidia’s chip. It can’t match Nvidia’s balance sheet.
Huang isn’t selling chips on credit. He’s buying loyalty with his credit rating.
What it defends against
Custom silicon. If the giants keep shifting to their own chips, Nvidia needs a second market that won’t. Financed labs and neoclouds are that market, and Kress’s 25% figure shows how big it’s getting.
The giants’ bargaining power. When a handful of customers buy a large share of your chips, they set the terms. A wider base of financed buyers gives Nvidia somewhere else to sell, and makes “we’ll build our own” a weaker threat at the negotiating table.
A demand air pocket. AI labs burn cash. If their funding dried up, orders would stop overnight. Nvidia’s guarantees keep the lights on through a funding gap and keep the orders coming.
The risk
The risk is the reason we’re cutting the score.
Kress’s defense was simple: “The hardware finds another buyer.” In a shortage, that’s true. Nvidia says demand still exceeds supply, so if one customer fails, another takes the chips. But guarantees don’t get called in a shortage. They get called in a glut, when GPU rental prices fall, borrowers can’t pay and the hardware is worth the least. The promise is safest exactly when it’s least needed.
The borrowers aren’t sturdy. CoreWeave, one of Nvidia’s key financed customers, reported about $72 billion in liabilities against about $5 billion in equity. Its quarterly interest bill more than doubled to $640 million. Nvidia’s own customers are taking longer to pay: days sales outstanding reportedly rose from 45 to 60. Nvidia has reportedly approached insurers about coverage for defaults and depreciation. Companies don’t buy that insurance for risks they think are remote.
There’s also the market’s verdict. When the $250 billion Ohio figure leaked, the stock fell 4.5% in a day, and the final guarantee came in at less than half of that. Investors are telling Huang exactly where their tolerance ends.
That’s why independence drops from 5 to 4. A year ago, Nvidia’s fate depended on what its customers wanted to buy. Now it also depends on whether they can pay, and a quarter of next year’s business comes from customers whose ability to pay rests partly on Nvidia’s own promise.
The lender and the borrower are now the same story.