Nvidia will guarantee the money OpenAI spends on Nvidia. Read who that makes the landlord.

There is a number in the reports about Nvidia and OpenAI that everyone is arguing about, and it is the wrong number. The number under debate is roughly 250 billion dollars — the financing guarantee Nvidia is reportedly in talks to provide so that OpenAI can lease a ten-gigawatt data-center campus in southern Ohio, with a separate discussion, on top of that, about financing as much as 350 billion in chips to fill it. The argument is whether that arrangement is circular, whether it inflates demand, whether it is the kind of thing that ends badly. Those are real questions and I will get to them. But they are the flattering questions — the ones that let you feel clever about accounting while the actual event goes by unremarked. The actual event is not a number. It is a transfer of terms. When the company that sells you the chips also guarantees the debt you take on to buy them, it has stopped being your supplier and started being your landlord.

Let me state the thesis plainly, because I would rather be wrong in the open than vague in the safe way. The reported backstop does not tell you that OpenAI is strong. It tells you the opposite, and then tells you who is strong instead. OpenAI — the most valuable private company most people can name, the one the entire industry treats as the protagonist — reportedly cannot raise this money on its own name. It does not carry an investment-grade credit rating. The debt markets, asked to lend against a ten-gigawatt bet, apparently wanted a co-signer. Nvidia can be that co-signer. That asymmetry is the whole story, and it survives every possible answer to the accounting question.

What is actually being built, and by whom

Start with the physical thing, because the physical thing is where power in this industry has quietly relocated. The project, as reported, is a campus on roughly 3,700 acres of federal land at Piketon, Ohio — the site of the former Portsmouth Gaseous Diffusion Plant, a Cold War uranium-enrichment facility now being repurposed into what would be, at ten gigawatts, the largest data-center project anyone has announced. SoftBank's energy arm is developing it. The first phase, around 800 megawatts, is reported for 2028. The all-in cost, chips included, is put at more than 500 billion dollars. And OpenAI's role in this — the detail that should reorganize how you read the whole thing — is tenant. It leases. It does not own the land, it does not own the power, and, if the chip-financing talks go the way they are described, it will not, in any meaningful sense, own the chips outright either.

So follow the compute, which is the only reliable way to find who is in charge in this business. The model is not the company; models leak and get cloned. The talent is not the company; it moves. I have written some version of that sentence for years. What I underweighted, and what this deal makes impossible to keep underweighting, is the layer beneath the compute: the credit that lets the compute get built. You can own the model and rent the chips. You can design the chips and rent the fab. And now, it turns out, you can be OpenAI and rent the balance sheet. Each time, the same thing happens — the party that can absorb the risk sets the terms for the party that cannot. The chips were already Nvidia's leverage. The financing is Nvidia's leverage compounded, because it converts a one-time sale into a standing claim.

The concession, made fairly, because it nearly is the story

Here is the counter-argument at its strongest, and it is strong enough that it dominated the headlines. Call it circular financing: a chipmaker lends its customer the money the customer uses to buy the chipmaker's chips, which lets the chipmaker book revenue and the customer book capacity, and the loop flatters everyone inside it. Michael Burry, who disclosed and expanded a short position in Nvidia, put it with a bumper-sticker's economy — around and around we go. Jim Chanos said the plainer version: Nvidia is effectively financing its own sales. The Bank for International Settlements, not a firm known for hot takes, flagged exactly these circular structures in its 2026 annual report as a source of systemic risk. That is a serious bench, and the worry is legitimate: a loop can manufacture the appearance of demand that a market would not have produced on its own.

And Nvidia has an answer, which deserves to be quoted rather than caricatured. The company reportedly sent analysts a seven-page memo rejecting the 'vendor financing' label outright — arguing its sales are ordinary arm's-length transactions, that it is typically paid within about 53 days, and that comparing this to the accounting frauds of the dot-com era, the Lucents and the Enrons, is a category error. Take that defense at full strength. Assume the revenue is clean, the demand is real, the payments are prompt, and no auditor will ever find anything. I am willing to grant all of it, because granting it is what exposes the point I actually care about.

The accounting question asks whether the money is real. The power question asks who wrote the terms. You can answer the first one entirely in Nvidia's favor and the second one still comes back Nvidia. — On what the memo can and can't settle

Because the accounting question and the power question are different questions, and the industry keeps letting the first one stand in for the second. The accounting question asks: is the revenue real? The power question asks: who set the terms under which it became real? Nvidia's memo, even if every word is true, is an answer to the first. It does not touch the second. In the cleanest, most legitimate, fraud-free version of this deal — the version Nvidia is describing — OpenAI still could not raise 250 billion dollars against its own credit, and Nvidia still could raise it against Nvidia's. That gap is not an accounting artifact you can memo away. It is the distribution of power in the AI economy, printed in a term sheet.

A lender that large is not a lender. It is a landlord.

What does it mean, concretely, to hold someone's terms? It means the guarantor's preferences quietly become the borrower's constraints. A company whose ten-gigawatt future depends on a specific counterparty continuing to guarantee its debt does not get to treat that counterparty like a vendor it can shop away from. It cannot credibly threaten to design the loop out — to move to a rival's accelerators, to a rival's financing — without threatening its own foundations. Nvidia has already put a reported 30 billion dollars of equity into OpenAI; Jensen Huang said in March that investment 'might be the last.' Layer a debt guarantee of this size on top of that equity, and you have a supplier that is simultaneously a shareholder and a creditor and the sole realistic source of the hardware. Three kinds of leverage, pointed the same way, at the same customer. The word 'customer' stops doing honest work at that point. A customer can leave.

This is the move I keep asking readers to watch for: the difference between power and the performance of power. OpenAI performs power beautifully. It sets the industry's agenda, it names the frontier, it gets the profile treatment and the policy meetings. Nvidia mostly performs modesty — a memo about payment terms, an investment that 'might be the last.' But performance is not the org chart. The party that can underwrite the other party's existence is holding the real thing, and it does not need to say so. It just books it, the way it books everything else.

And there is a second landlord in this building

Now widen the frame one more click, because there is a detail in the reporting that most coverage buried and that changes the geometry entirely. The power for this campus — the electrons, ten gigawatts of them — is reported to be controlled by the U.S. government and funded separately by Japan under a recent trade deal, with Commerce Secretary Howard Lutnick involved in deciding which companies get access to it. Read that slowly. The chips and the credit belong, in effect, to Nvidia. The power belongs to Washington, financed by Tokyo. Stack the layers of this supposedly-OpenAI project and you get something closer to a sublease three landlords deep:

• The land is federal — a decommissioned government uranium site, leased out.
• The electricity is a government-rationed resource, paid for under a bilateral trade arrangement, with a cabinet secretary deciding who is allowed to draw it.
• The financing is guaranteed by the chip supplier, because the tenant's own credit will not carry it.
• The chips are the supplier's, potentially financed by the supplier too.
• What OpenAI brings to the arrangement is the demand, the brand, and the operating story.

That is not a criticism of OpenAI's engineering, which is real, or its demand, which is also real. It is a description of dependency. The company at the center of the AI story turns out to be the one part of this structure that owns the least of it. Everyone else in the stack holds an asset that is scarce and hard to replicate — federal land, rationed power, guaranteed credit, the only chips that matter. OpenAI holds the part that is, in the long run, the most contestable: the lead in a model race where the leader changes every few months.

Where I was wrong

I owe the reader a correction, because a column that never revises itself is just a brand. For a long time I assumed that a market with many hungry buyers would discipline a dominant seller — that so many companies wanting chips would keep Nvidia honest, that demand this broad would diffuse power rather than concentrate it. I had the vector backwards. When the buyers are this dependent and the seller is this singular, breadth of demand does not check the seller's power; it deepens it. Every additional buyer that cannot fund its own buildout is another party that needs Nvidia to co-sign. The dependency runs from the many to the one, and the reported OpenAI backstop is simply the largest, clearest instance of a direction the whole market has been travelling. I was too sanguine about the crowd. The crowd is the leverage.

Let me grant the most generous reading one last time, because it is also the most unsettling. Suppose the defenders are right — suppose Dario Amodei and the investors at firms like Janus Henderson are correct that this is simply how you finance a genuinely explosive, genuinely real build-out, that the demand is authentic and the structure is rational. I think there is a lot to that. But notice what it concedes. If the demand is real and durable, then the terms written around it are real and durable too. A bubble unwinds and frees everyone trapped inside it. A sound, permanent structure does not unwind — it hardens. The circular-financing skeptics are worried this ends in a crash. The more interesting worry is that it doesn't: that it works exactly as designed, and the arrangement in which one company guarantees the money its customers spend on its own products becomes the permanent architecture of the most important industry of the decade.

So here is the question the 250-billion-dollar headline is designed to keep you from asking. Not 'is the money real,' which Nvidia can answer, and not 'is it circular,' which the short-sellers can argue. The question is: when the terms of the most important company in artificial intelligence are set in a contract between its chip supplier and a government that rations the power, who actually decides what OpenAI does next? It is not a trick question, and the uncomfortable part is how short the answer is. Not OpenAI. Follow the compute down to the credit that builds it, and the company you find holding the terms is the one that also sells the chips — which is to say, the landlord was never going to be the tenant with the most famous name. It was always going to be whoever could sign the lease.