On a credit desk, the question has changed. It is no longer whether Nvidia can sell accelerators, but who absorbs the loss if the companies buying them cannot pay for them.
Five-year credit-default swaps on Nvidia’s debt finished Oct. 27 at 127 basis points, according to data compiled by S&P Global Market Intelligence. That is roughly double the level of the CDX North America investment-grade index and higher than the six largest U.S. banks in the same data set. In plain terms, the bond market now charges more to insure Nvidia’s obligations than to insure the institutions that lend to it.
The repricing has been quick. The same contracts cost 39 basis points at the end of 2025, which puts the widening at about 226%. Few large technology issuers have moved that far in a year without an earnings shock. Nvidia has not had one.
On Oct. 21 the spread pushed above Meta Platforms’ for the first time since November 2025. The ordering usually runs the other way. A supplier with pricing power and a customer that depends on it should be the safer credit, and for most of the past two years that is exactly how the market priced the pair. Meta is Nvidia’s largest customer. It is now also, in the eyes of swap traders, the better risk.
What is being priced is not distress. At 127 basis points Nvidia remains a comfortable investment-grade name, and its cash generation would cover its obligations many times over in any ordinary year. What is being priced is role. Nvidia now does several things chip companies historically did not. It holds a stake in CoreWeave. It has committed to buy computing capacity that the cloud provider cannot sell. It provides partial guarantees on data-center leases signed by partners.
Each of those obligations is disclosed, bounded and manageable on its own. Stacked together, they make Nvidia the backstop for a slice of the AI buildout that runs on rented silicon. If demand for rented compute falls, the first losses land on the operator, the second on the landlord, and the third on the company that guaranteed the lease and promised to buy the idle capacity.
The market’s reference point is CoreWeave itself. Its five-year swaps have sat above 600 basis points for most of this year, a level that usually belongs to borrowers with a live question about repayment. The distance between 127 and 600 is, roughly, the price of Nvidia’s balance sheet standing between those leases and the people who own them. Traders are not saying Nvidia resembles CoreWeave. They are saying Nvidia has taken on a piece of CoreWeave’s risk.
Analysts said the move reflects concentration rather than borrowing. The distinction matters for how the spread behaves from here. A company that is merely indebted pays down and the spread tightens. A company that has become a guarantor tightens only when the guarantees lapse or the underlying assets prove liquid. Nvidia has given no timetable for either.
The same day the swap data circulated, Nvidia said it would subscribe to $3 billion of new Meta shares, widening cooperation on its Rubin platform, Vera CPUs, Spectrum-X networking and Muse agents. A company buying its customer’s stock in the same week that its own credit is being repriced against that customer’s is a compact illustration of how the AI supply chain now finances itself. The equity purchase is an investment. It is also, functionally, a subsidy to demand.
There is precedent, and it is not encouraging. Equipment suppliers lent to telecom carriers in the late 1990s, booked the orders as revenue, and then watched the carriers fail with the equipment still on their books. Analysts who cover the current cycle have drawn the comparison repeatedly, usually to explain why they discount vendor-financed revenue.
The counterargument is that the current buildout is funded by companies with genuine cash flow, that the guarantees are partial, and that Nvidia’s exposure is a rounding error against its gross profit. That is true today. It was also true of the equipment vendors in 1999, in the specific sense that the commitments were small relative to the profit they supported.
What would narrow the spread is unglamorous: clearer disclosure of the guarantee stack, conversion of capacity commitments into contracted revenue, and the retirement of backstops as they expire. What would widen it is another round of them, particularly if the next one is larger and attached to a customer with less cash than Meta.
For now the swap market has assigned Nvidia a cost of protection that implies something the equity market has not: that the company’s largest risk sits on the other side of its own sales, in the balance sheets of the customers it has helped finance, and that the price of insuring against it is now higher than the price of insuring the banks that arrange the deals.
