Nvidia Is Selling Its Credit Rating

Read the $250 billion number reported this weekend and the interesting part isn't the size. It's what the money doesn't buy. Nvidia's proposed backstop for OpenAI's Ohio campus covers the lease and the debt on the building. Not the chips. The chips get financed separately, in a different conversation, for up to another $350 billion. Which means the thing Nvidia is putting on the table first isn't silicon at all. It's a balance sheet. The company stopped being purely a chip vendor somewhere in the last few quarters and started renting out the strongest credit in technology so its customer can qualify to buy.

Here's the shape of it, per the Journal's Sunday report. SoftBank is developing a 10-gigawatt campus in southern Ohio — Piketon, on the bones of a former uranium enrichment plant. First 800-megawatt phase targeted for 2028. Fully built, the project runs north of $500 billion including the hardware inside it, which would make it the largest data center ever announced. Nvidia is in talks to guarantee roughly $250 billion of the lease and the debt behind it. Talks are early and could fall apart. Michael Burry summarized the structure in five words: around and around we go.

The tell is the credit rating, not the dollar figure

Why does a company with OpenAI's revenue need someone to co-sign?

Because it isn't investment grade. That's the load-bearing fact in the whole story, and it's buried three paragraphs down in most of the coverage. Debt at this scale prices off the borrower's rating, and the gap between investment grade and not, on $250 billion of twenty-year paper, is measured in tens of billions of dollars of interest. OpenAI can't close that gap on its own credit. So Nvidia lends its own.

Strip it to first principles and ask what actually got transacted. Not compute. Not chips. Creditworthiness — moved from the company that has it to the company that needs it, in exchange for the purchase order at the other end. Nvidia's product line now includes counterparty risk, and unlike the GPUs, that line doesn't show up as revenue. It shows up as a contingent liability in the footnotes, if it shows up at all.

Lucent already ran this experiment

This is where I get genuinely interested, because we have the tape from last time.

In the late nineties, Lucent Technologies had a demand problem that looked exactly like an opportunity. The competitive local exchange carriers wanted to buy switching gear and couldn't fund it, so Lucent financed the purchases itself. On the order of $8.1 billion committed to customers who existed largely because Lucent was willing to write the check. WinStar alone got a $2 billion commitment. Booked as sales. Recognized as revenue. Celebrated as growth.

Then the capital markets closed. WinStar went under in April 2001 owing Lucent something past $800 million, and Lucent wrote off roughly $700 million on that one name. Provisions for bad customer debt: $2.2 billion in 2001, another $1.3 billion in 2002. Three and a half billion dollars of revenue that had already been reported, celebrated, and priced into the stock, reversing out the back end as credit losses.

The lesson people took was "vendor financing is bad," which is the wrong lesson and too simple. Vendor financing isn't fraud and it isn't even unusual. What it does is subtler and worse: it converts a sales problem into a credit problem and then hides the credit problem inside a revenue line. Lucent didn't lie about anything. It just recognized income today for risk it kept on its own books, and the two numbers didn't reconcile until the cycle turned.

The diversification is fake

Now zoom out to the thing everyone tracks. Hyperscaler capex for 2026 sits somewhere around $750 billion, spread across five or six names, and it reads as a broad distributed bet. Different companies, different balance sheets, different risk. That's the mental model, and it's the reassuring one.

Guarantees collapse it. If the marginal deal at the frontier of the buildout is underwritten by one company, then the apparent diversification across six borrowers is actually concentration behind one guarantor. Correlation you can't see on the capex chart goes to one the moment demand disappoints. And the guarantor is the same firm booking the revenue when the deal closes. That isn't a conspiracy. It's just a loop, and loops don't have a natural place to absorb a loss.

The honest counterargument matters here, so take it seriously. Lucent's customers had no revenue and no demand. OpenAI has both, and every hyperscaler on the last earnings cycle said the same thing: supply-constrained, not demand-constrained. If that holds, the guarantee never gets called and this is simply how you finance infrastructure with a long payback and a short depreciation clock. Fine. But notice the shape of the bet. Vendor financing doesn't create demand. It pulls demand forward. And the mechanism that pulls it forward is precisely the mechanism that concentrates the damage if the demand shows up late.

One more detail that deserves more attention than it's getting: the power for the Piketon site is controlled by the federal government, funded by Japan under a trade agreement, with the Commerce Secretary weighing in on who gets it. Compute allocation just became industrial policy. That's a separate post, but file it — the scarce input isn't chips anymore, and it isn't capital either.

Stop watching the capex number. Start watching how much of it is guaranteed by someone who gets paid when the deal closes.

A backstop is just revenue with a delay and a footnote. The footnote is where you find out who was actually buying.

— Dustin