Hi,
Nvidia has put billions behind the companies that use its chips. It has bought stakes in AI labs such as OpenAI and Anthropic, promised to cover part of the bill on data centers built for them, and agreed to rent computing time back from cloud companies such as CoreWeave that buy its hardware.
Critics call it circular. With Jensen Huang gone, a sharper question emerges: who keeps financing the AI buildout? Huang was willing to make enormous bets on the companies buying his chips. His successors, and the investors funding those customers, may not be.
So we traced the money.
How much is exposed
In August, Colette Kress, then CFO and now interim chief executive, told investors that demand from the AI labs Nvidia backs would make up roughly a quarter of its business next year. On the $695 billion analysts expect for Nvidia's next fiscal year, that is about $170 billion.
That sounds alarming. It is less than it looks. Most of those chips are bought by Microsoft, Oracle and other clouds with their own money, then rented to the labs. We traced the $170 billion through disclosed contracts and project plans, counting each chip once, at the company that buys it.

About $35 billion sits in projects whose money or buildings are not settled. That is about 5% of next year's forecast. It is three projects, and about $26 billion of it serves one lab, Anthropic.
Where the risk actually sits
Take Nscale, which is building a West Virginia campus for Anthropic. It plans to spend $21 billion to $22 billion on chips and other equipment for the site, and its November 12 filing says some of its new orders cannot be deferred, reduced or cancelled. The same filing says it has no binding commitments yet for the financing tied to its Anthropic contracts: those financings "remain in process."
Equipment it must pay for, without the money yet raised to pay for it. If that financing fails, someone takes the loss. But Nscale buys Nvidia's chips through server makers, not from Nvidia, and the filing does not list Nvidia as a lender. The unsettled money is mostly other people's.
Why Nvidia can wait and its customers can't
Nvidia calls next year "a supply-constrained outlook." Its buyers say the same. Microsoft: "Customer demand continues to exceed available supply." CoreWeave: ten times as much qualified demand as capacity through the end of 2027.
And it has happened before. In 2025 Microsoft walked away from about 2 gigawatts of leases, and Meta and Google were reported to take them. In March 2026, after Oracle and OpenAI dropped a planned expansion in Abilene, Texas, Microsoft took the 900 megawatt campus within about three weeks.
While chips are scarce, a project that can't raise its money frees hardware someone else is waiting for. Nvidia can sell a scarce chip to the next buyer in line. A data center developer can't hand off its debt.
The catch is the building, not the chip. A shipment can move between buyers in weeks. A new powered data center takes 12 to 15 months, and vacancy is about 1%. Most of these chips will find a buyer. The question is whether that buyer has the money and a building ready to plug them in.
Our Tessara research piece this week puts a number on what a financing failure would cost Nvidia, explains why that protection weakens and why we see 2028 as the period to watch, and sets out the two dated forecasts we have put on our public record.
Tell me what you think we're missing.
Teng
P.S. Want to follow the money yourself? Ask Tessara:
If the AI labs' financing slips, which buyers have powered buildings ready for Nvidia's chips in 2027?
