Data-Center and AI Infrastructure Debt Sells Off as Execution Risks Reprice

A broad sell-off in debt tied to data centers and AI infrastructure is widening beyond newer cloud AI providers to project-backed securities associated with large technology tenants.

CoreWeave bonds fell below 80 cents on the dollar, with yields at one point on Thursday approaching 15%. Project debt has also weakened across several data-center developments: QTS bonds for Project Excalibur in Georgia, where Microsoft is the tenant, fell below 88 cents on the dollar. Bonds issued by Beignet Investor for the Hyperion project traded at about 92 cents, down from above par in early July.

Hut 8 project bonds traded at about 94 cents on the dollar for its Louisiana project and about 92 cents for its Texas project, which is leased by NVIDIA. Construction bonds for Oracle's AI park in Michigan declined from par to 87 cents on the dollar.

The pressure is also affecting financing tied to Oracle's Project Jupiter in New Mexico. The associated loan is about $18 billion, according to S1, and has faced pressure amid delays in power-supply permitting. The available information does not specify the duration of those delays, the final status of the permits, or their effect on the project schedule.

The market move is being linked to questions about the economics of AI capital expenditure, as well as power availability, permitting readiness and construction execution. Local political resistance to new US data centers—citing concerns over water use, noise and air pollution—adds a further development risk for projects that still require approvals.

Gartner forecasts that 80% of new cloud companies will cease operating independently by 2030. It estimates that AI data-center operators would need to generate cumulative revenue of $8.5 trillion by the end of 2030 to cover capital and financing costs, while spending on consumer and enterprise AI services flowing to data-center providers would not exceed $3.6 trillion. Gartner also warns that, if the current investment trajectory persists, more than $4 trillion of assets could face impairment risk by 2032.

For data-center financing, the episode may signal a shift away from underwriting based primarily on anticipated AI demand and tenant strength toward closer differentiation by grid access, permitting progress, construction delivery and cash-flow resilience. This could translate into more expensive and selective debt for new projects before commissioning. Already built and operational data centers may be relatively more valuable if constraints restrict new supply, although that view is a market participant assessment cited by S1 rather than an established outcome.