Nvidia Stopped Buying Into OpenAI. It Started Renting Out Its Credit Rating.
The Wall Street Journal reported on Sunday, and Bloomberg confirmed on Monday, that Nvidia is in talks to guarantee roughly $250 billion of financing behind OpenAI's planned data center campus in Piketon, Ohio. Ten gigawatts, developed by SoftBank's energy arm, first phase around 800 megawatts targeted for 2028, total project cost plausibly north of half a trillion dollars once you count the silicon. Separately, Nvidia is discussing financing as much as $350 billion of the chips that go inside it. Bloomberg put the full basket of arrangements above $750 billion.
Most of the coverage filed this under circular financing, which is the same headline we have been running on Nvidia deals since last year and is not wrong. It is just not specific. The number changed, sure. The instrument changed too, and that is the part worth your attention.
A Guarantee Is Not an Investment
When we wrote about Nvidia crossing $40 billion in AI equity commitments back in May, every one of those deals had the same shape: cash leaves Nvidia, an asset arrives on the balance sheet, and the worst case is that the asset goes to zero. Painful, bounded, disclosed. You can look up what Nvidia paid for its OpenAI stake and know the maximum loss.
A guarantee does not work that way. No cash moves on signing. Nothing shows up as an expense. Nvidia does not own anything new. What it has done instead is promise lenders that if OpenAI cannot service the debt on a building it is leasing, Nvidia will. The exposure is contingent, it is off the balance sheet until it is triggered, and the ceiling on it is whatever the underlying obligation grows to. That is a fundamentally different risk object than a stock certificate.
| Instrument | Cash out day one | Where it sits | Maximum loss |
|---|---|---|---|
| Equity stake | Full amount | Assets, marked | What you paid |
| Vendor loan | Full amount | Receivable | Principal plus interest |
| Guarantee | Zero | Footnote, until triggered | The whole obligation |
Read the last row again. The instrument with the largest downside is the one that costs nothing today and lives in a footnote. That combination is not a scandal by itself, plenty of ordinary corporate finance runs on guarantees, but it does mean the usual quick checks people apply to these deals stop working. You cannot spot it in cash flow from investing. It does not dent free cash flow. It will not compress reported margins. If you were tracking Nvidia's customer exposure by watching what it spends, you just went blind.
The Credit Market Was Not Blind
Here is the number that made this a story rather than another Sunday leak. Nvidia's five-year credit default swap, the contract that pays out if Nvidia itself defaults, jumped to a record 82 basis points on Monday. That was the biggest single-day move since the contract started trading actively in November 2025. The stock fell about 6 percent across two sessions, which got the headlines, but equity moves for a dozen reasons on any given Monday. CDS moves for one.
Nvidia's default protection now costs more than Alphabet's. Sit with that. This is a company that did $215.9 billion of revenue in fiscal 2026, printed $81.6 billion in a single quarter, generated $50.3 billion of operating cash flow in that quarter, and carries gross margins near 75 percent. On the fundamentals nothing about Nvidia got riskier this week. What got riskier is the set of promises attached to it.
| Exposure | Reported size | Form |
|---|---|---|
| Ohio campus lease and debt | ~$250B | Guarantee, in talks |
| Chip purchase financing | Up to ~$350B | Financing, in talks |
| Prior OpenAI commitment | ~$30B | Equity, trimmed from $100B |
| Other 2026 AI equity | $40B+ | Equity, dozens of names |
Nothing in the top two rows is signed. Terms are unsettled and the whole package could shrink or die, the way the original $100 billion OpenAI number quietly became something closer to $30 billion. But the CDS market does not wait for a signature to reprice a tail.
Why a Guarantee at All
This is the question that answers itself, and the answer is the actual news.
OpenAI has no investment-grade credit rating. On the numbers most widely circulated it is losing something near $14 billion this year against roughly $25 billion of revenue. A lender being asked to fund a multi-decade lease on a 10 gigawatt campus in southern Ohio, against a tenant with those financials and a product cycle measured in months, is being asked to underwrite a very long asset with a very short view of the borrower.
The guarantee exists because that underwriting did not clear on its own. Nvidia's balance sheet is being substituted for a credit rating OpenAI does not have, so the paper can be priced against Nvidia's risk instead of OpenAI's. That is the entire mechanism. Which means the existence of the guarantee is itself a disclosure about what the debt market thinks of financing OpenAI unassisted, and it is not flattering.
I want to be fair about the counterargument, because there is a real one. Nvidia is not guaranteeing a speculative venture with no demand behind it. It is guaranteeing capacity for the largest consumer AI product on earth, and if the tokens keep flowing the building gets paid for out of revenue that already exists in some form. Structurally this is closer to a landlord taking a corporate guarantee from a parent company than to a startup writing IOUs. Guarantees mostly do not get called.
The problem is correlation. The scenario where OpenAI cannot pay its Ohio lease is not an isolated tenant problem. It is a scenario where AI demand disappointed badly enough to break the biggest name in the category, which is the same scenario where Nvidia's order book collapses, its equity stakes in a dozen AI companies mark down together, and it would very much prefer not to be writing a check on someone else's building. The guarantee pays out precisely when Nvidia can least afford it. That is what the CDS is pricing, and the CDS is right to price it.
Our Take
The circular financing frame has been load-bearing for a year and it is starting to sag. Calling equity stakes, warrants, prepayments, and now credit guarantees all the same thing because money moves in a loop tells you almost nothing. What matters is who holds the downside and how it is disclosed, and on both counts a guarantee is the most aggressive structure anyone in this cycle has proposed.
It also marks a shift in what Nvidia is selling. Compute vendors used to compete on chips, then on rack-scale systems, then on software stacks, and now the frontier of the competition is the ability to make a customer bankable. AMD put $5 billion into Anthropic in milestone-gated cash with no warrants, and we called that the cleanest deal in the market. That view holds up better every week. Nvidia is playing a different game with a much bigger balance sheet, and the price of the game just showed up in its own credit spread.
Three things worth watching. Whether any of this gets signed at the reported size or shrinks on the way to a term sheet, since the $100 billion to $30 billion precedent suggests leaks run hot. Whether Nvidia discloses guarantee exposure as a named line item in its next 10-Q or leaves it in commitments and contingencies, because that choice tells you how the company itself is classifying the risk. And whether the CDS holds above the pre-leak level after the initial reaction fades, which is the cleanest available read on whether credit investors think this is a headline or a repricing.
Nvidia spent 2026 turning its cash into stakes in its own customers. The Ohio deal is a different trade. It is lending out the one asset that cannot be diversified, and the market charged it for the privilege the same afternoon.