US AI Data-Center Debt Reprices as Credit Markets Focus on Execution Risk
Редакция ЦОД:Главное
Bonds tied to several US AI data-center projects have traded below par, despite links to major technology tenants. CoreWeave securities faced the sharpest pressure, with its bonds falling below 75 cents on the dollar and the yield on a 2031 maturity approaching 16%, according to S1.
- Bonds linked to multiple US AI data-center projects were reported to be trading below par.
- The reported weakness extends to projects associated with major technology tenants, including Meta, Microsoft, NVIDIA and Oracle-linked development.
- CoreWeave bonds fell below 75 cents on the dollar, while a 2031 maturity’s yield at one point approached 16%.
- Power interconnection, permitting, construction timing and cost control may become more important in AI data-center debt underwriting.
- CoreWeave reported a $104.2 billion order backlog at the end of the second quarter, plus more than $25 billion of additional commitments early in the third quarter.
US credit markets are applying greater scrutiny to debt-financed AI data-center development, with bonds associated with several projects trading below par, according to S1. The pressure spans projects linked to large technology tenants and highlights a shift in lender focus from tenant credit quality alone to project execution risks.
S1 cited Hyperion bonds issued by Beignet Investor, involving Blue Owl and Meta, trading at about 91 cents on the dollar. Bonds for Hut 8 projects in Louisiana and Texas were cited at about 93 cents and 91 cents, respectively. QTS Project Excalibur in Georgia, leased by Microsoft, was reported below 87 cents on the dollar.
The source also identified pressure on a roughly $18 billion loan associated with Oracle-linked Project Jupiter. S1 attributed this to delayed power approvals amid local opposition. The case illustrates the importance credit markets may place on the certainty and timing of grid interconnection, alongside permitting and construction execution.
CoreWeave has seen the most pronounced market stress among the examples cited. Its bonds fell below 75 cents on the dollar, while the yield on its 2031 notes at one point approached 16%, according to S1. The source reported that CoreWeave has about $50 billion of debt and raised its annual CAPEX forecast by $4 billion in August to $35 billion-$39 billion. It also cited expected annual revenue of close to $13 billion and an expected annual loss of about $2.13 billion.
CoreWeave has pointed to contracted demand: S1 reported a $104.2 billion order backlog at the end of the second quarter and more than $25 billion of additional commitments early in the third quarter. Those figures represent company statements and do not independently establish debt-service capacity.
At the sector level, S1 cited Gartner estimates that AI companies could invest nearly $1 trillion in infrastructure in 2026 and about $6 trillion more between 2027 and 2030. Gartner’s estimates, as reported by S1, indicate that AI data-center owners would need around $8.5 trillion in cumulative revenue by the end of 2030 to cover CAPEX, financing and normal returns, while AI revenue attributable to the data-center layer may not exceed $3.6 trillion.
The market repricing could move AI data-center financing toward a more rigorous assessment of power availability, permitting readiness, construction schedules and project-level debt-service capacity. Models combining large upfront CAPEX, high leverage and delayed revenue appear most exposed to that tighter underwriting approach.
The repricing suggests that a blue-chip tenant alone may no longer offset lenders’ concerns over power interconnection, permitting, construction timing, cost escalation and the pace at which AI demand converts into revenue. This could raise the cost of debt and make financing more selective, particularly for highly leveraged operators with large upfront CAPEX and deferred cash flow.
- — / 16 октября 2026 г. · КитайскийПервоисточник →
Материал подготовлен редакцией ЦОД: Главное на основе указанных первоисточников.