OpenAI's survival is now someone else's credit risk

On July 9, S&P Global Ratings cut Oracle to BBB-/A-3 from BBB/A-2, one notch above junk. The reason the agency gave was not a product or a quarter or a balance sheet. It was a customer. In its rationale, as relayed through secondary reporting, S&P treated OpenAI as a central credit risk and warned that if OpenAI couldn’t meet its payment obligations, Oracle would be left holding long-term datacenter rental agreements. It couldn’t easily re-let on comparable terms. A rating agency looked at Oracle and priced in the possibility that one AI lab runs out of money.

For two years the fear ran one direction. Everyone was building on the labs, so if a lab went dark, everything above it broke. I made that case in “Too big to fail, again”: the pipeline goes to zero rather than degrading, and the headcount that used to be the fallback is gone. Take it as settled. What the S&P action marks is the inverse. The unprofitable labs have become the largest customers of the companies renting them compute, and the dependency now runs up the capital stack, to the lessors who signed the leases and the bondholders who funded them. Everyone still depends on the lab. Now its landlord’s creditors are watching it too.

This isn’t the loop described when Nvidia financed its own buyers. That was seller money on both sides of demand: the chip vendor funding the customers who buy the chips, so the demand and the capital came from the same place. This is linear, not circular. No one here is funding their own revenue. A landlord signed a long lease, rented the space to a tenant who might not survive the term, and the credit market repriced the landlord’s debt accordingly. Ordinary counterparty risk, except the counterparty is a company burning tens of billions a year, and the party now formally pricing the risk is a rating agency.

The mechanism is a duration mismatch. Oracle has disclosed roughly $248 billion in datacenter lease commitments running 15 to 19 years. The tenants renting that capacity sign much shorter terms. The Oracle-OpenAI arrangement under Stargate runs about $300 billion over roughly five years; CoreWeave’s OpenAI commitments total around $22.4 billion through 2029, also about five. The neocloud version is sharper, and Jeffrey Moerdler of Haynes Boone put it plainly to Bisnow: “They’re signing 15-year data center leases, and their customer-facing agreements are by the hour, the day, the month, max one to three years.” He called neoclouds “WeWork 2.0.” The landlord holds the long-dated liability. The tenant holds the short-dated commitment, and the tenant holds the losses. OpenAI’s cash burn is projected around $27 billion in 2026, a figure that is estimated and in places leaked, not audited. Fifteen to nineteen years of fixed obligation on one side, five or fewer on the other, and the party on the short side isn’t making money.

The market is already charging for the gap, and the cleanest evidence is a single piece of paper. In May, CoreWeave’s term loan drew about $15 billion in orders, roughly 4.8 times oversubscribed, and the spread tightened. By early August the same facility repriced about 125 basis points wider, and CoreWeave accepted maintenance covenants that had been, in the lenders’ own framing, absent from most major leveraged loan deals for over a decade. Same instrument, one quarter apart. The bull case and the bear case are the same debt, and the market changed its mind in ninety days.

It’s not one instrument alone. Oracle’s five-year credit default swaps hit a record around 200 basis points, and traders described them explicitly as a liquid hedge on OpenAI execution: not on Oracle’s products, but on whether its largest tenant performs. The rating agencies have converged. Moody’s cut Oracle’s outlook to negative back in September 2025, citing a high reliance on revenue from a single counterparty, and in a sector report dated July 24 or 25 it coined the phrase “circular AI ecosystem,” naming Oracle and CoreWeave as the weakest links. Oracle’s own 10-K says the quiet part in filing language: if customers don’t renew, it may be “unable to re-lease, repurpose or assign such capacity on acceptable terms, if at all,” and it notes that OpenAI’s ability to pay “depends entirely on its ability to continue raising capital.”

There are good arguments against this. The lease contracts are take-or-pay. OpenAI’s own services agreement makes its minimum commitments non-cancellable and, on a non-cause termination, immediately due. The tenant can’t simply walk. CoreWeave booked roughly 96 percent of its 2024 revenue under take-or-pay terms. Capacity is the binding constraint: about 98 percent of Oracle’s AI capacity is already contracted, which means a failed lab’s space plausibly re-lets into a shortage rather than sitting empty. The hyperscalers behind much of this retain strong balance sheets. And the sharpest cut against a clean story comes from Moody’s itself, which notes that hyperscalers keep their leases short and off-book by disclaiming renewal certainty. So, the tidy framing of a 15-to-19-year lease against a five-year contract doesn’t hold uniformly across the field.

It holds where the ratings moved, at Oracle and the neoclouds. This isn’t a prophecy of collapse. The take-or-pay clauses are precisely why a rating cut, not a default, is what happened.

Lenders have changed who they worry about, and they are charging for it. For two years the risk pointed down, from the labs to everything built on them. It now points up, to whoever holds a long lease against a short-dated, unprofitable tenant: Oracle, the neoclouds, and the bondholders behind both. On July 9 a rating agency wrote one AI lab’s solvency into another company’s credit line, and set it one notch above junk.