The Wall Street Journal reported Sunday that Nvidia is discussing roughly $250 billion in financing guarantees to let OpenAI lease a 10-gigawatt campus that SoftBank’s energy arm is building in southern Ohio. Nobody involved has confirmed it. The sourcing is unnamed people familiar with the matter, which at this stage of a deal usually means the talks are real and the terms are not settled.
The headline number is the least interesting part. What the guarantee covers is the lease and the construction debt — not the chips. The chips are a separate conversation, reportedly worth up to another $350 billion in vendor financing, and the whole thing lands somewhere north of half a trillion dollars once you count the silicon going into the building.
So Nvidia would be underwriting the building, financing the hardware, and selling the hardware. Three roles, one balance sheet.
The reason any of this is necessary is straightforward. OpenAI has no investment-grade rating. Ask a lender to underwrite twenty years of lease payments from a company whose revenue history is shorter than the construction timeline and whose cost base is dominated by the equipment being installed, and the paper either prices badly or doesn’t clear. Nvidia’s credit fixes that by substitution — lenders stop pricing OpenAI and start pricing Nvidia. The guarantee stays contingent, off the primary balance sheet, invisible until it isn’t.
Michael Burry got to the obvious objection within hours: the chip vendor is guaranteeing the money that buys the chips. The defence is that these are separable instruments and that the guarantee attaches to real estate and power infrastructure, which hold residual value regardless of who occupies them. That argument works only if the campus has a second tenant at similar economics in the world where OpenAI can’t pay. A 10-gigawatt facility built to one customer’s spec, defaulting in a scenario that almost by definition involves the whole AI compute market repricing at once, is not an asset with a deep bid. The telecom equipment vendors made a version of this argument in 1999.
The Part That Isn’t About Money
Buried in the reporting is the detail that actually determines whether this happens. The project’s power allocation is controlled by the US government and financed separately by Japan under a recent trade arrangement, with Commerce Secretary Howard Lutnick involved in deciding who gets it. OpenAI has been in advanced talks for weeks and is described as among the keenest bidders. Anthropic, Microsoft and Google have all been in the room too.
That reorders everything. Capital is not scarce — AI infrastructure spending clears $700 billion this year. Chips are constrained but they can be bought. Firm multi-gigawatt power, available on a timeline measured in years rather than decades, is the input nobody can manufacture their way around, and here the allocation sits with a cabinet officer rather than an interconnection queue.
Which reframes the $250 billion. It’s a financing structure, yes. It’s also a bid for a permit, and a way of handing the government the easiest decision to defend.
What This Would Mean If It Closes
The precedent isn’t the size. It’s chip vendors becoming credit intermediaries for the infrastructure that consumes their product. Every one of Nvidia’s three positions here is individually defensible — suppliers extend terms, strategic investors take exposure, guarantors get paid for risk. Stacked, they mean a slowdown in AI compute demand hits revenue, receivables and contingent liabilities in the same quarter. That’s not diversification. That’s the same bet placed three ways.
Plenty of ways for this to die quietly. Rating agencies could treat the guarantee as debt-equivalent, which removes the entire point of structuring it as a guarantee. The board could balk at contingent exposure running to a meaningful slice of market cap. Documentation could stall on whether the lease guarantee and the chip financing are genuinely separate instruments with separate triggers, or one package wearing two names — and that question is worth more attention than any dollar figure in the reporting, because the answer decides whether the circularity criticism sticks.
Watch the credit market rather than the equity market on this one. Equities are bad at pricing contingent guarantees; they mostly ignore them until they convert. Nvidia’s CDS spreads will tell you whether lenders think the credit substitution is real or cosmetic. Shares closed Friday at $206.84, down under a percent, before any of this circulated.
Also worth watching: whether the power allocation goes to one tenant or gets split. A split award would say the government is thinking about concentration risk in national compute capacity, and that changes the arithmetic for everyone who bid.
Leave a Reply