Analysis
The AI race stopped being just a model race months ago, and the clearest evidence is Nvidia's own balance-sheet strategy: in August, Nvidia partnered with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize more than $500 billion in third-party capital for AI infrastructure, according to Fortune. Our read: access to cheap, large-scale financing is now as decisive a competitive edge as any model benchmark, and the companies built to win on compute architecture alone are the ones most exposed if credit tightens.
The numbers back the shift. The five largest U.S. hyperscalers are on pace to spend $730 billion on capital expenditure in 2026, financing up to a third of it with debt, per Reuters figures cited in Fortune's reporting. Nvidia's financing platform effectively tries to turn GPU compute into an asset class that pension funds and insurers can own the way they already own infrastructure and real estate -- which only works if those investors believe AI demand is as durable as a toll road or a power plant.
“hyperscalers are on pace to spend $730 billion on capital expenditure in 2026, financing up to a third of it with debt, per Reuters figures cited in Fortune's reporting.”
That's where the U.S. holds a real structural edge over rivals: the Stanford HAI AI Index estimates U.S. private AI investment last year ran more than 23 times China's, and eurozone households sit on roughly $11 trillion in bank deposits that European policymakers are only now trying to mobilize toward AI infrastructure through financial reform rather than new technology. China, meanwhile, is reportedly sitting on computing capacity it can't use -- state media put the share of idle newly built capacity at 80%, per MIT Technology Review.
Room for disagreement: this financing wave assumes AI demand keeps compounding fast enough to service $500 billion-plus in new obligations, and that GPUs don't depreciate fast enough to leave today's hardware as stranded collateral in three years -- a risk pension funds and insurers aren't traditionally built to underwrite. Pulse has already tracked junior-tranche yields on AI infrastructure debt climbing from 8.5% on Project Big Sky to above 10% on Anthropic's Project North Star, with CoreWeave's bonds trading near 15% -- a cost of capital that's rising even as Nvidia tries to make compute look safer to own, not riskier.
Whoever wins the next phase of the AI race will have spent as much energy on term sheets with Apollo and Blackstone as on training runs -- and competitors without Nvidia's balance sheet or Wall Street's attention will have to build their own financing machinery from scratch, or get boxed out by whoever already has.