AI Debt Costs Rise as Data Center Expansion Gets Pricier

The financing math behind the AI buildout is getting tougher. Treasury yields have climbed to their highest levels since 2007, raising borrowing costs for the data center companies racing to add capacity. CNBC reports that JPMorgan estimates $4.1 trillion in AI-related debt could be issued through 2030.

For now, higher rates have not stopped the spending spree. But they are changing which projects can attract capital, and how much risk lenders are willing to accept.

Neoclouds face the sharpest pressure

The biggest companies in the AI ecosystem can generally borrow more cheaply because they have investment-grade credit ratings and diversified cash flows. Smaller GPU cloud providers and data center operators have less room to absorb the increase. CNBC cited market participants who expect financing to become more selective, even when borrowers offer higher returns.

CoreWeave illustrates the exposure. The company has warned that each one-percentage-point increase in rates could add about $30 million to its annual interest expense, based on its floating-rate debt. SoftBank, meanwhile, raised $11.1 billion in junk bonds this week, with yields reaching 9.75% on part of the offering.

Why demand is still winning

AI companies are signing long-term compute commitments, giving infrastructure operators a reason to keep building even as capital gets more expensive. OpenAI and Anthropic are among the model developers locking in capacity, while hyperscalers continue to commit hundreds of billions of dollars to capital expenditure.

That demand makes this less like a normal slowdown and more like a sorting mechanism. Projects with credible customers, reliable power, and strong utilization may still get funded. Speculative capacity, or facilities dependent on optimistic forecasts, will have a harder time clearing the financing hurdle.

Builder impact: AI builders should treat infrastructure economics as part of platform risk. Higher financing costs can affect GPU availability, cloud pricing, reserved-capacity commitments, and the viability of smaller providers. Design workloads to move between providers, avoid hard-coding a single capacity assumption, and measure the cost of idle or over-reserved compute before committing to long contracts.