A quarter-trillion-dollar signature

Here's the deal: on July 27, 2026, the Wall Street Journal reported that Nvidia is in talks to provide a financial guarantee worth roughly $250 billion on OpenAI's behalf.

"Guarantee" is doing a lot of work in that sentence, so let's unpack it. OpenAI wants to sign a 20-year lease on an enormous data center campus in southern Ohio. Lenders financing that campus looked at OpenAI's credit and got nervous. So the proposal is that Nvidia promises to cover the obligation if OpenAI can't. It's a co-signer arrangement — structurally identical to a parent co-signing an apartment lease, except the number has twelve digits.

And that's not the whole package. The same reporting says Nvidia is separately discussing around $350 billion in financing for the chips that would fill the campus. Add them up and you get roughly $600 billion of Nvidia-linked commitments attached to one project. Total project cost is estimated at more than $500 billion, which would make it the largest single data center build announced anywhere.

If something feels off about that, your instincts are fine. Nvidia is the company selling chips into this campus. It would also be underwriting the money to buy those chips, and the rent on the building that houses them. The seller is financing the buyer. That's the story.

The four parties at the table

Nvidia dominates AI accelerators, and its constraint has quietly shifted. The bottleneck is no longer "can we make enough chips" or "is our architecture better." It's "can our customers raise the money." Demand exists in enormous quantities; the balance sheets behind that demand do not always keep up. Nvidia already has an investment partnership with OpenAI reported at up to $100 billion. This guarantee pushes that relationship into new territory.

OpenAI built the most recognized product in consumer AI and still isn't profitable. That's the hinge of this whole story. Private valuation doesn't matter to a project-finance lender; cash flow and credit rating do. OpenAI does not carry an investment-grade rating, which means it cannot, on its own signature, commit to a two-decade lease of this magnitude.

SoftBank and SB Energy are the landlord side. SB Energy, SoftBank's energy subsidiary, is developing the campus — acquiring the site, securing power, building the shells. Developers borrow tens of billions against future rent, and that borrowing only works if lenders believe the tenant will still be paying in year 18.

The lenders are the real gatekeepers here. Their skepticism about OpenAI's credit is precisely what created the need for Nvidia. A guarantee from a highly rated counterparty lowers the interest rate and extends the tenor of the debt. Same building, same tenant, dramatically different cost of capital.

What's actually on the table

Item Detail
Nvidia guarantee ~$250B (real estate and lease obligations)
Separate chip financing ~$350B under discussion
Total project cost $500B+ estimated
Site Piketon, Ohio — former Portsmouth uranium enrichment facility
Developer SB Energy (SoftBank subsidiary)
Campus capacity 10 gigawatts
Lease term 20 years
Phase 1 800 MW, targeting 2028
Status In negotiation — no signed contract

The site deserves a paragraph of its own. This land in southern Ohio housed a Cold War-era uranium enrichment plant, and that history is the entire reason it's attractive. Facilities like that were engineered to consume staggering amounts of electricity, so the transmission infrastructure is already in the ground. The parcel is large, the neighbors are few, and the grid interconnection exists. This is why data center developers now spend their time studying decommissioned steel mills, retired power plants, and old nuclear sites. The scarce resource in AI is no longer silicon. It's electricity.

Ten gigawatts is hard to picture, so anchor it: a typical nuclear reactor produces somewhere around one gigawatt. This campus at full build would draw the output of roughly ten reactors. That's peak summer demand for a major metropolitan area, consumed by a single computing campus. The first phase is more modest at 800 MW, aimed at 2028.

One caveat worth repeating: nothing is signed. The WSJ described talks. Neither Nvidia nor OpenAI has confirmed terms publicly. Deals at this scale routinely get restructured or abandoned.

Who gets what

Nvidia gets demand certainty. Its largest risk right now isn't a competitor's chip — it's whether customers can keep funding purchases. Backstopping this lease effectively reserves accelerator volume for a 10-gigawatt campus over two decades. The accounting is also convenient: a guarantee is a contingent liability. No cash leaves until something goes wrong. Revenue lands now; the exposure sits in the footnotes, conditional on a future that Nvidia is betting won't arrive.

OpenAI gets time and scale. Terms it could never access on its own credit become available. A 20-year lease is a commitment to occupy this campus into the 2040s — and locking in compute that far out is a defensive move in a race where the constraint is capacity, not ideas.

SoftBank gets an anchor tenant. Masayoshi Son has bet heavily on AI infrastructure, and infrastructure development lives or dies on who signs the long lease. Landing the most recognized AI company for 20 years, with an Nvidia guarantee stapled to it, transforms the financing profile of the asset.

Lenders get credit substitution. They're no longer underwriting OpenAI risk; they're underwriting Nvidia risk. For a credit committee, that's a completely different conversation.

Now the uncomfortable part. This structure doesn't eliminate risk. It relocates it. If OpenAI can't generate enough revenue to cover two decades of rent, the obligation migrates to Nvidia's balance sheet. But the scenario in which that happens is a scenario where AI demand has broadly collapsed — which is the same scenario that guts Nvidia's core business. One company would be absorbing two risks that fail together. That correlation is the weakest joint in the whole arrangement.

How this pattern has played out before

Sellers financing buyers is called vendor financing, and it has a track record that cuts both ways.

The cautionary tale is late-1990s telecom equipment. Lucent Technologies and Nortel sold gear to newly formed carriers and lent them the money to pay for it. Revenue exploded. So did the stock. Then the dot-com bust took out the carriers, and the receivables turned into write-offs. Lucent absorbed massive charges and never recovered as an independent company. Revenue booked on the way up came back as losses on the way down — that's the textbook version of this failure mode.

But the pattern also works. Aircraft manufacturers have financed and leased planes to airlines for decades, and that market functions normally today. The difference is collateral quality. An aircraft is standardized; if one airline fails, the plane gets re-leased to another. Residual value survives. Large real estate is similar — the building outlasts the tenant, and another tenant can be found.

Apply that test here and the answer splits. A permitted 10-gigawatt site with existing transmission is genuinely scarce. Even in a demand correction, that land and those interconnection rights would attract someone. That profile looks more like aircraft leasing than like Lucent. But hundreds of thousands of a specific GPU generation depreciate fast. Which is why the $350B chip financing is the riskier half of this package, not the $250B real estate guarantee — even though the real estate number is the headline.

There's also the recurring circularity critique. Nvidia invests in customers, customers buy Nvidia chips, those purchases show up as Nvidia revenue. Skeptics have hammered this loop for a while; Nvidia has consistently argued each transaction stands on its own commercial merit. This guarantee will pour fuel on that argument, because it goes a step beyond equity investment — it takes on someone else's contractual obligation.

How rivals respond

Google doesn't need this. Alphabet funds data centers from operating cash flow and runs its own TPUs. The news is free positioning for them: we build without co-signers. They are, however, competing for the same megawatts.

Microsoft and Amazon sit in between. Both hold top-tier credit ratings and finance capacity without help, and both push in-house silicon (Maia, Trainium). For them the signal here is about land and power, not money. The number of US sites that can deliver gigawatt-scale interconnection on any reasonable timeline is small, and it's shrinking.

AMD faces a different problem. It has been closing ground with large AI supply agreements, but if Nvidia starts deploying its balance sheet as a competitive weapon, the contest shifts from performance-per-watt to financial capacity. That's a battlefield that favors the far larger company. Expect AMD to lean harder into the "we compete on merit, not on financing loops" message.

Anthropic and other frontier labs face the same compute math. If OpenAI locks in a $600 billion financing package, everyone else needs a comparable patron or a fundamentally different strategy. The visible trend is already labs binding themselves tightly to specific chip suppliers and clouds.

Then there's regulation and power allocation, which is not a footnote. Directing gigawatts to individual projects has become a federal policy question in the US, with Commerce overseeing aspects of allocation. Local resistance is growing too, as residents in multiple states link data center buildouts to rising electricity bills. Financing is one gate. Permits, interconnection queues, and politics are others.

What actually changes

If you invest, add a line item to your Nvidia analysis: contingent liabilities and customer financing exposure. Guarantees typically live in the notes rather than on the face of the balance sheet, and at this scale the notes are worth reading directly. Also: this is a report about negotiations, not a signed deal.

If you run an AI startup, the takeaway is compute bifurcation. When the largest players pre-commit decades of power and land, everyone else buys leftover capacity at spot cloud pricing. If you can't sign a 20-year offtake, efficiency and specialization beat scale as a strategy — not as a consolation prize, but as the only rational play.

If you work in energy or infrastructure, this is straightforwardly good news. Brownfield industrial sites, idle transmission capacity, and generation projects keep appreciating. Data center development has become a power business wearing a real estate costume.

If you follow policy, watch concentration of systemic risk. If one company's balance sheet underwrites credit across the AI buildout, that company's stress becomes the sector's stress. Regulators have spent years worrying about exactly this shape of interconnection in finance. No framework currently governs it here.

If you're just following the story, keep one sentence: the AI race moved from models to infrastructure, and now from infrastructure to finance. Who builds the better model matters less this quarter than who can raise half a trillion dollars.

🥄 Three Things You're Probably Wondering

— Could Nvidia actually be on the hook for $250 billion? In principle, yes — but only in a world where OpenAI can't pay rent, which is a world where AI demand has broken down. It also matters that the guarantee sits behind real estate, so there's a recoverable asset underneath; it isn't a straight write-off. The real problem in that scenario is that Nvidia's core business would be deteriorating at the same time.

— Isn't this just circular financing? The critique is clean: Nvidia funds the buyer, the buyer buys Nvidia chips, that becomes Nvidia revenue. The rebuttal is also reasonable: a guarantee isn't cash out the door, and the leased asset is a building, not GPUs. Which reading holds up depends on accounting treatment that hasn't been disclosed. Calling it either way right now is premature.

— Will it really be running in 2028? That's the target for the 800 MW first phase. Large data centers slip most often on grid interconnection and substation equipment, and this scale makes that worse. The site's existing power infrastructure genuinely helps. Still, megaprojects that hit their announced dates are the exception, not the rule.

Sources

Numbers and terms are as reported and not yet contractually final, so they may change. Investment calls are yours to make!