Nvidia 洽谈为 Nscale 西弗吉尼亚 Anthropic 数据中心提供债务担保
Core Highlights
According to a Thursday report by The Information, later corroborated by Bloomberg, Nvidia is in talks to provide billions of dollars in debt credit support for a West Virginia data center that AI cloud provider Nscale is planning. In plain terms, the chip giant is considering backing a customer's loan: if Nscale cannot repay construction costs, Nvidia would temporarily take over the facility and help find a new operator. It is the first time the company has extended support from leasing and chip purchases into data center construction itself.
What Happened
Nscale's IPO filing discloses that Anthropic has signed a contract worth as much as 44.6 billion dollars for the campus's first building, expected to use Nvidia's next-generation Vera Rubin chips with 460 megawatts of capacity. A second building worth roughly 45 billion is also under discussion. Nscale itself has troubles: in August only about 25,000 advanced AI chips were running, far short of a deployment target near 500,000; its West Virginia site has repeatedly flooded nearby homes with construction runoff; and Microsoft withdrew from a planned UK site this month. Nscale is seeking to list at a valuation of up to 35 billion dollars.
Technical Details
The arrangement is known as a residual value guarantee. Nvidia's previous support for cloud providers mainly covered leasing and chip purchases, not civil construction. This guarantee is designed to help Nscale secure power and building financing. According to The Information, Nvidia has already provided roughly 170 billion dollars in equity injections and credit support to new cloud providers and major customers, including a one-billion-dollar convertible financing commitment to Nscale in September.
Comparison and Impact
CoreWeave, Nebius and IREN all saw their shares fall more than 30 percent in October, reflecting investor unease with a model that has little cash-generating core business and relies on rental income to service debt. The more deeply Nvidia binds these customers, the more it locks in chip demand — but it also exposes its own balance sheet to infrastructure cycle risk.
Further Context
For the AI infrastructure chain, a chipmaker stepping in to backstop customer debt signals that capital's tolerance for enormous capex is approaching a limit. Anthropic filed for its own IPO this week, yet a "junk" rating from agencies could raise its financing costs. This story is a key signal for watching the resilience of the AI build-out: when the chain of guarantees extends back to the chipmaker itself, the path along which a collapse would propagate becomes shorter and more direct.
There is a historical echo here. Vendor financing — a supplier lending money to customers so they can buy the supplier's products — has appeared in earlier technology booms, from telecom equipment in the late 1990s to various hardware cycles. It can sustain demand longer than fundamentals alone would justify, and it concentrates risk exactly where the industry can least afford it. That does not mean the arrangement is doomed; data centers are durable assets with real demand. It does mean observers should watch the structure closely, because the same mechanism that accelerates growth can amplify a downturn. For investors, the practical question is not whether AI demand is real, but whether the financing structures built to meet it can survive a period of slower revenue realization.
Additional Notes
The IPO dimension adds another layer of complexity. Nscale is seeking to list at a valuation of up to 35 billion dollars while its deployed chip count lags far behind plan and one of its anchor tenants has signed a contract whose scale strains credibility. Public-market investors will scrutinize exactly the metrics the guarantee is meant to shore up. A residual value guarantee can make a financing look safer than the underlying cash flows justify, and sophisticated buyers know to trace who ultimately bears the risk. If the structure is perceived as circular — a chipmaker backing a customer that buys the chipmaker's chips — the valuation may face pressure regardless of the headline contracts.
The construction problems are not incidental either. Repeated flooding of nearby homes from site runoff points to execution risk on the physical side, which is the part of AI infrastructure least forgiving of shortcuts. Power, water, permitting and local relations are as decisive as silicon, and delays there can push out revenue recognition by quarters. Microsoft's withdrawal from a planned UK site illustrates how quickly a marquee partner can change course.
For industry watchers, the useful takeaway is to read such announcements as a stress test of the entire financing chain. The question is never whether demand for compute exists — it plainly does — but whether the elaborate structures built to fund it can withstand a slower ramp or a repricing of risk. When the chipmaker itself becomes the backstop, the system has fewer degrees of freedom than it appears.