Nvidia sells chips. On Sunday the Wall Street Journal reported it is now negotiating to do something else entirely: guarantee roughly $250 billion in financing so OpenAI can lease a 10-gigawatt data center built on a decommissioned Cold War uranium enrichment site in southern Ohio. The guarantee covers the lease and the debt behind the buildings. It does not cover the chips. For those, Nvidia is separately discussing financing OpenAI’s purchases of up to another $350 billion in its own hardware.
Read that structure twice, because it is the whole story. The company selling the GPUs is prepared to co-sign the loan for the real estate that houses them, and then lend the buyer the money to buy the GPUs. Reuters could not independently verify the WSJ report, negotiations are described as early, and terms could collapse or change. But the direction is not ambiguous, and it is not new. It is the largest, most literal version yet of a pattern the AI industry has spent 2026 normalizing.
What the deal actually is
The site is the former Portsmouth Gaseous Diffusion Plant near Piketon, Ohio, roughly 3,700 acres of federal land that spent the Cold War enriching uranium. It is being rebranded the PORTS Technology Campus. SoftBank’s energy arm, SB Energy, is the developer. At full buildout the campus is designed for 10 gigawatts of compute and up to 10 gigawatts of new power generation, most of it (around 9.2 gigawatts) from natural gas built on-site, with AEP Ohio handling roughly $4.2 billion in grid and transmission upgrades. The power itself sits on government-controlled land and is funded separately by Japan, tied to Tokyo’s pledge under a US-Japan trade deal to put about $33 billion into the natural gas plant.
The full campus could cost at least $500 billion at current prices for silicon, construction, and power. OpenAI would control the computing equipment under a 20-year lease and, critically, would not begin paying until the site starts operating. The first phase, roughly 800 megawatts, is targeted for 2028. The first gigawatt is meant to run on Nvidia’s Vera Rubin platform.
So the money moves like this. Japan funds the power. SB Energy develops the shell. Nvidia guarantees OpenAI’s ability to make the lease payments and guarantees SB Energy’s project financing. Nvidia then also finances OpenAI’s purchase of Nvidia chips. OpenAI, the tenant, pays nothing until 2028 and currently loses about $1.22 for every dollar of revenue it books. Conventional debt markets looked at that income statement and declined to fund the gap on their own. That refusal is the reason Nvidia is in the room.
The vendor became the lender of last resort
I have spent 20-plus years on the buyer side of infrastructure in telecom, and there is a bright line in how you evaluate a vendor’s financing help. Vendor financing to smooth a purchase (net-90 terms, a leasing arm, a deferred first payment) is ordinary and healthy. What the Piketon structure describes is different in kind. Nvidia is not smoothing a purchase. It is underwriting the counterparty’s solvency across a 20-year horizon so that the counterparty can keep buying from Nvidia. When the guarantor of the loan and the seller of the collateral are the same company, the transaction has stopped being a sale. It has become a bet the seller is placing on its own demand.
This is the same shape the industry has been drawing all year, each iteration larger than the last. Google put $40 billion into Anthropic while selling it TPUs and cloud. Yesterday we covered AMD taking equity in Anthropic so Anthropic will buy AMD chips. Nvidia’s own October 2025 arrangement with OpenAI involved up to $100 billion of staged investment as capacity came online. The Piketon guarantee is a different beast only in magnitude and mechanism: it is not equity, it is a credit backstop, which means Nvidia carries the downside without necessarily holding a matching claim on the upside. If OpenAI cannot pay the lease in 2029, the guarantor pays. The chip maker has quietly moved from the top of the capital stack to the bottom.
The constraint moved from chips to power and balance sheets
For two years the binding limit on AI was silicon. You could not get enough H100s, then you could not get enough Blackwell, and the whole industry organized itself around Nvidia’s allocation list. That era is closing. Look at where the Piketon money and the friction actually are: 9.2 gigawatts of new gas generation, a $4.2 billion transmission upgrade, a Japanese sovereign commitment to build the power plant, a federal land transfer, and a decade-long construction timeline that Counterpoint Research’s Neil Shah flatly called out (“a 10-gigawatt site won’t just appear overnight and will take at least a decade to fully build out”). The chips are almost an afterthought in that list. They are the one component you can actually procure on a normal schedule.
The scarce inputs now are power, permitted land, transmission, and creditworthy balance sheets. That is why a former uranium enrichment plant is valuable: it comes with grid interconnection, federal cooperation, and enough acreage to matter. It is why SoftBank chose France for an $87 billion build on the strength of its nuclear supply, why NextEra’s $67 billion utility merger was structured around AI load, and why the interesting number at Piketon is 9.2 gigawatts of gas, not the count of Vera Rubin racks. Compute has become a physical-infrastructure business wearing a software company’s margins, and the two do not fit together without someone bridging the gap. Right now that someone is Nvidia.
Why this concentrates risk instead of spreading it
The optimistic reading is that these “symbiotic” deals, as Shah put it, are simply how capital-intensive infrastructure gets financed when the tenant is young and the asset is long-lived. Railroads and telecom fiber were built on comparable leaps of faith. And OpenAI’s revenue, while dwarfed by its commitments, is real and growing fast.
The problem is concentration. Nvidia now sits on three sides of the AI economy at once: it sells the chips, it invests in or lends to the buyers, and at Piketon it guarantees the buildings those chips sit in. Greyhound Research’s Sanchit Vir Gogia noted that in these arrangements “the relationship stops being vendor and customer,” and scarcity itself “becomes contractual” through minimum commitments and reservation tiers. When one company is supplier, financier, and landlord-guarantor to the same customer, a stumble by that customer does not stay contained. It hits Nvidia’s product revenue, its investment book, and its guarantee obligations simultaneously. Diversification is supposed to be the thing that makes a system survivable. This is the opposite: the same balance sheet backstopping the same demand it created.
OpenAI’s own CFO, Sarah Friar, reportedly flagged in April that the company had signed roughly $1.4 trillion in infrastructure commitments that achievable revenue might not support. OpenAI has since reset its investor framing toward roughly $600 billion in compute spend through 2030 against a target of around $280 billion in 2030 revenue, up from about $13 billion last year. Those are enormous, and enormously uncertain, numbers. We walked through the same gap in detail when OpenAI filed its trillion-dollar IPO paperwork showing it loses $1.22 per dollar of revenue. The Piketon guarantee is what it looks like when a company that unprofitable still needs to sign a 20-year lease: someone else has to promise the landlord it will get paid.
What a buyer should take from this
If you are evaluating AI vendors, the practitioner lesson is not “the bubble is popping,” a claim I would not make from a WSJ report on early-stage talks. It is narrower and more useful. First, the AI supply chain now has a single dominant point of failure, and it is not a chip shortage, it is Nvidia’s willingness and ability to keep underwriting its own customers. That is worth watching more than any benchmark. Nvidia grew revenue 85% last year after losing all of China, so the capacity to backstop is real today; the question is what happens to the whole structure if that growth so much as flattens.
Second, keep your own architecture portable. The deals being signed above your head are locking specific tenants to specific silicon on specific sites for two decades. Your defense is the same as it was six months ago: qualify a second model, keep your integration layer vendor-neutral, and do not let a favorable price today mortgage your optionality in 2028. The companies signing these guarantees are betting the demand curve only goes up. Your job is to stay solvent and flexible if it doesn’t.
The uranium site is a fitting stage for all of it. The last time the federal government poured this kind of money into Piketon, it was building the fuel for a Cold War it hoped never to fight. This time the arsenal is compute, and the guarantor is the company that makes the ammunition.