Investors are increasingly selling off debt tied to data centres supporting artificial intelligence (AI) infrastructure amid growing concerns that revenues from the AI sector may not be sufficient to cover the large financing costs. Bonds linked to major US data centre projects, where companies such as Meta, Nvidia, and Oracle are key tenants, have recently traded significantly below their face value.
The debt instruments backing flagship projects like “Hyperion” and “Project Excalibur” have seen their prices decline sharply. For instance, bonds issued by the Meta-backed special purpose vehicle Beignet Investor for the Hyperion project fell from above par value in July to just under 92 cents on the dollar. Similarly, debt for data centres developed by Hut 8 in Louisiana and Texas—leased by Nvidia—has dropped to roughly 94 and 92 cents on the dollar. Bonds for Blackstone-owned QTS’s Project Excalibur in Georgia, leased to Microsoft, have decreased below 88 cents on the dollar.
The situation is more severe for so-called "neocloud" providers, which have been aggressively financing AI infrastructure growth. Nvidia-backed CoreWeave bonds have plunged as low as 75 cents on the dollar, a level often viewed as distressed, pushing yields above 15 percent. CoreWeave’s stock declined by 22 percent over the past week, while fellow provider Nebius lost 20 percent. Nvidia itself—a major investor in multiple neoclouds and AI startups—has seen its market value contract by more than $300 billion within the same time frame.
Market skepticism has been fueled by developments such as OpenAI’s delay of its initial public offering until next year and slower sales growth at Anthropic, prompting questions about the sustainability of the AI boom. Analysts caution that there is a substantial gap between the capital expenditures pouring into AI-related infrastructure and the revenue streams these investments are expected to generate.
Ludovic Subran, chief investment officer at Allianz, described the current environment as a “Darwin moment” for data centre bonds, where the weakest credits face survival challenges amid expanded credit spreads caused by power shortages and construction delays. This suggests that the current pricing underestimates execution risks involved in these projects.
A recent report by technology consultancy Gartner predicts that 80 percent of neocloud operators may cease independent operations by 2030 due to a significant imbalance between capital spending and expected returns. Gartner estimated that owners of AI data centres will require $8.5 trillion in revenues by the end of 2030 to cover capital and financing expenses and provide attractive returns. However, the consultancy forecasts that consumer and corporate expenditures on AI services flowing to data centre providers will likely reach no more than $3.6 trillion over the same period, leaving a multitrillion-dollar shortfall.
Gartner warned that if current investment trends continue, over $4 trillion in AI infrastructure asset values could face writedown risks through 2032. This growing caution among investors is underscoring the challenges of sustaining the rapid expansion of AI infrastructure projects amid uncertain financial returns.