Read Nvidia's own Q1 FY2027 10-Q and you'd conclude the company barely does this. Maximum gross exposure across every partner lease guarantee it has disclosed: $3.5 billion, shrinking as partners pay down their leases. Fair value: "immaterial." Structure: credit derivatives, taken in exchange for warrants, backed by $712 million sitting in escrow. This is the language of a company that has convinced its auditors the whole program is a rounding error.
Now read the Wall Street Journal. Nvidia is in talks to guarantee roughly $250 billion behind OpenAI's lease of a 10-gigawatt campus in Piketon, Ohio.
That is 71 times the entire disclosed guarantee book. More than a year of Nvidia's revenue. About four times the $62.6 billion in cash, equivalents, and marketable securities the company held when fiscal 2026 closed.
That is 71 times the entire disclosed guarantee book—more than a year of Nvidia's revenue, about four times the cash it held when fiscal 2026 closed.
The two numbers describe the same activity. Only one of them is honest about what it now is.
This week the escalation is cleaner and stranger: Nvidia is no longer just selling the chips or investing in the customer. It is co-signing the customer's real estate.
The temptation is to file this under vendor financing scaled up. Nvidia has done versions of it before—$30 billion into OpenAI, backing for CoreWeave, a reportedly up to $500 billion lease of a Hut 8 facility in Texas. The lease guarantees looked like the mild end of that spectrum. So a $250 billion guarantee reads as the same thing, just with more zeros.
It isn't. It's a category error, and the reason lives in the remedies.
Go back to the first guarantee, disclosed in Q3 FY2026. It was capped at $860 million, with $470 million of escrow behind it—reportedly more than half the exposure pre-funded in cash. That alone made the "immaterial" label defensible.
But the escrow wasn't the real protection. The real protection was optionality. The partner had separately contracted to sell the data center's cloud capacity, so a second revenue stream stood between Nvidia and a loss. And if both the escrow and that contract came up short, Nvidia retained the right to assume the lease for its own internal use, or to sublease it.
That is what made the guarantee genuinely low-risk. Nvidia was underwriting a building it could actually want. If the tenant vanished, Nvidia could walk in, plug in its own GPUs, and run the thing—or hand it to someone who would. The downside had a floor because the asset was fungible and Nvidia was a natural buyer of its own.
Now hold that structure up against Piketon, and watch every piece fail to survive the trip.
Start with the asset. This is the former Portsmouth Gaseous Diffusion Plant, a site that enriched uranium for the U.S. weapons program from 1954 until 2001 and is still being decontaminated. SB Energy leases the land from the federal government; it doesn't own it. Nvidia cannot "assume the lease for internal use" on federal land it has no standing to occupy, at a scale—10 gigawatts—it has no operational need to absorb.
The sublease remedy is equally dead. To sublease 10 GW of capacity you need another tenant who wants 10 GW at Piketon. The candidates are Anthropic, Microsoft, and Google—and all three have spoken to Commerce Secretary Howard Lutnick, who controls the site's power allocation, about it in recent weeks. Nvidia would not be subleasing into a captive market. It would be handing capacity to the very rivals it's trying to keep buying its chips, in a negotiation it doesn't control.
Then the escrow. At $860 million of exposure, $470 million in escrow was 55% coverage. At $250 billion, matching that ratio would require roughly $137 billion parked in cash—more than twice everything Nvidia has. There is no world in which a Piketon guarantee is meaningfully pre-funded. The escrow cushion that made the small deals safe simply cannot scale.
And the power itself is not a settled input. The campus needs 9.2 gigawatts of new natural gas generation plus $4.2 billion of transmission work with AEP Ohio, financed by $33.3 billion Japan committed under its trade deal with Washington. The first phase—roughly 800 MW—isn't expected until 2028. Nvidia would be guaranteeing lease payments on capacity whose electricity depends on gas plants not yet built and a federal official's allocation decision not yet made.
Every load-bearing element of the original design—assume, sublease, escrow, a fungible asset, a captive resale market—is gone. What remains is a naked guarantee against a bespoke, immovable, half-decontaminated site whose power supply is a political variable.
That is why the $3.5 billion book and the $250 billion number are not the same instrument at different sizes. The small guarantees were structured so real remedies capped the risk. The Piketon guarantee has no remedies, which means the risk is genuinely uncapped. Carrying it as an "immaterial" credit derivative would be an accounting fiction stretched past the point of description.
Now the circularity, which is the part the market started pricing this week.
OpenAI has no investment-grade credit rating. Nvidia's involvement reportedly reflects that gap. SB Energy can't raise $250 billion in debt against OpenAI's balance sheet, so it raises the debt against Nvidia's instead. Nvidia lends OpenAI the creditworthiness OpenAI lacks.
OpenAI then uses that borrowed creditworthiness to lease the building—so it can fill the building with Nvidia chips. And Nvidia is separately discussing financing those chips too, a deal reportedly worth another $350 billion.
So trace the loop. Nvidia guarantees the real estate so OpenAI can rent space to house Nvidia GPUs that Nvidia also finances, to serve compute demand from a company Nvidia has already put $30 billion into. A single balance sheet stands behind the building and its contents, the tenant's credit and the equipment's financing.
What Nvidia is doing here is converting demand risk into counterparty risk on itself. The old worry was whether customers would keep buying chips. The new structure answers that worry by guaranteeing the customer can pay—using Nvidia's own credit as collateral. The demand looks more certain precisely because Nvidia has underwritten it. That is not the same as the demand being real.
The commenters on the trade press put it more bluntly than any analyst will on the record: OpenAI can't get a bank to co-sign the lease for the data centers it needs to sell a product it currently sells at a loss, so Nvidia is considering co-signing a quarter-trillion in obligations that may never be repaid. Write off the $30 billion already invested and Nvidia survives. A $250 billion guarantee that goes bad is a different order of event.
None of this is signed. Terms aren't settled, and the WSJ notes the arrangement could still collapse. Nvidia may yet find a structure with real remedies, or walk away.
But the disclosure gap is the story regardless of outcome. A company whose filings describe its guarantee program as a $3.5 billion afterthought is negotiating a single guarantee 71 times that size, against an asset none of its safeguards can reach. The tidy fiction in the 10-Q was true when the guarantees were small and the remedies were real. At Piketon, both conditions break at once—and the number the market has to reckon with is the one Nvidia's own filings were built to make look immaterial.