The market that financed the AI buildout is beginning to ask whether it lent too much. Roughly $18 billion in loans tied to a data-center project that Oracle leases in New Mexico is trading under pressure, with underwriters including Santander and Jefferies quoting prices at 89 to 91 cents on the dollar, according to a Financial Times report. The buyers’ worry is twofold: the credit standing of Oracle itself, and local opposition to the water and pipeline permits the project needs.
The project carries the code name Project Jupiter and spans 1,400 acres, a scale that makes it one of the larger data-center plays in a state that has become an unlikely destination for the AI boom. Lenders extended credit on the assumption that the site would move from dirt to running servers on schedule. The discount now priced into the loans says the market is no longer certain of that assumption.
The concerns are specific and political. Data centers are thirsty, and the project’s need for water and for the pipelines that move it has run into opposition from local interests who question whether a giant server farm should draw on resources in a region that guards them closely. Permitting fights of this kind can stall a project for years, and a stalled project is precisely what the discount is pricing.
There is now more than market chatter. Separate reports said the project has been temporarily shelved, a development that, if it holds, would transform the loans from a financing of construction into a bet on a site that may not be built as planned. Around the same period, Oracle’s flagship Stargate data-center operation was reported to be cutting staff.
Reuters had already flagged the slide in prices on September 18. A week later, the notable fact is not that the loans traded down but that the discount has not repaired. When debt trades at 89 cents and stays there, the market is not expressing a temporary wobble; it is demanding a risk premium for the possibility that the collateral never becomes what was promised.
The episode is a window into how the AI infrastructure boom has been financed. The equity side of the story has been euphoric: companies promising enormous data-center capacity have seen their shares bid up on the prospect of a computing buildout that has no clear ceiling. The debt side has been quieter, a pile of loans and bonds extended to projects whose cash flows lie years in the future.
Debt, unlike equity, does not have the luxury of waiting on a story. Loans must be repaid on a schedule, and their value is a hard judgment about whether the cash will be there. When the judgment sours, it shows up first in the price of the paper, which is why the debt side of the AI build is feeling the cold before the stock side has had to acknowledge it.
Oracle’s own credit remains the other variable. The company has committed to enormous capital spending to fund its data-center expansion, and investors have been willing to finance that ambition at rates that assumed its balance sheet could absorb the strain. A loan trading at 89 cents is, in part, a referendum on whether that assumption still holds.
The company has said little about the New Mexico project’s status, and the reports of a pause have not been confirmed publicly. What is confirmed is the price: underwriting banks are quoting a discount that would have been unthinkable when the loan was arranged, and the discount has held for a week.
Analysts said the situation bears watching not because one project matters, but because it may be the first visible crack in a financing structure that has been stretched across the entire industry. If lenders begin to mark down data-center credits broadly, the cost of capital for the buildout rises, and the boom slows by arithmetic rather than by choice.
Oracle’s ambitions in AI have been among the loudest in the industry. The company has said it plans to spend heavily on data centers to meet demand for computing, and its chairman has spoken repeatedly of the scale he intends to reach. The New Mexico site is one piece of that plan, and the loans against it are a direct test of whether lenders believe the spending will produce the cash flows it promises. For the moment, the discount says the jury is still out.
For the moment, the exposure is concentrated. One project, one state, one set of permits, one loan trading at 89 cents. But the AI buildout was never going to be tested by its successes. It was always going to be tested by the first project that stalled, and by how the lenders who financed it reacted when the payments stopped looking certain.
