AI's Debt Binge Gets More Expensive as Bond Yields Hit Near 20-Year Highs

Key points

• JPMorgan Chase estimated in June 2026 that $4.1 trillion in AI-related debt will be issued through 2030.
• The 10-year US Treasury yield reached approximately 5.17% this week, its highest level since 2007, up about one percentage point since the start of 2026.
• SoftBank raised $11.1 billion in a junk-bond sale this week, paying yields as high as 9.75% on its seven-year debt.
• CoreWeave's SEC filing warns that a one-percentage-point rise in rates adds roughly $30 million to its annual interest bill.
• Lenders are narrowing the field: one finance executive told CNBC that the market is truly interested in only around 20 of the roughly 50 neocloud companies, the smaller specialist cloud firms built specifically to run AI workloads.

Building the data centres that run AI costs an almost incomprehensible amount of money. The industry has been borrowing at a pace that makes other sectors look cautious. Now those loans are getting noticeably more expensive.

The 10-year US Treasury yield, a benchmark that sets the floor for what companies pay when they borrow money, climbed to around 5.17% this week. Highest since 2007. Every company refinancing existing debt or raising new money is looking at a more painful bill.

Who gets hurt most?

Big tech can absorb it. Amazon, Google, Meta and Microsoft all carry investment-grade credit ratings, which means lenders charge them less because the risk of default is low. They've committed to hundreds of billions in capital spending this year and are expected to increase that figure in 2027.

For everyone else, the calculus is harder.

Neoclouds, specialist cloud companies built to run AI workloads, often carry more debt relative to their size and have less cushion to absorb rising costs. A senior private credit investor, speaking anonymously to CNBC, said those deals are becoming harder to finance. Riley Thompson, a vice president at Mitsubishi HC Capital America, put a number on it: instead of funding a roster of 50 neoclouds, the market is genuinely interested in perhaps 20. AI2Day first covered the neocloud financing market on 4 August 2026, when two-year-old British firm Nscale was already chasing $3.5 billion ahead of a possible IPO.

CoreWeave, a neocloud that listed on the Nasdaq stock exchange earlier this year, spells out the risk plainly in its latest SEC quarterly filing. As of June 2026, a one-percentage-point rate increase would add about $30 million to its annual interest costs on floating-rate debt, meaning debt whose rate moves up and down with the market.

SoftBank, which borrowed $11.1 billion to close its $30 billion OpenAI bet, sold those bonds at yields as high as 9.75%, junk-bond territory that pays higher rates to attract buyers willing to take more risk. Mark Malek, chief investment officer at Siebert Financial, told CNBC the company had little choice: "They basically are price insensitive to that raise, which means they're price takers."

Will demand slow the borrowing down?

Probably not.

Meta's Muse personal assistant app clocked more than 2.5 million global downloads in its first two weeks after launching in September 2026, passing ChatGPT at the top of Apple's App Store. Demand like that makes it hard to pump the brakes. Andrew Giudici, global head of infrastructure finance at credit rating agency KBRA, told CNBC he expects large debt issuance to continue regardless of rate levels.

Bernie Margulies, CEO of AI infrastructure advisory firm American Compute, summed up the logic bluntly. Companies with supply contracts already signed with OpenAI or Anthropic, contracts that lock in compute capacity years ahead, aren't going to walk away over half a percentage point. "If you have a deal with Anthropic, will 50 basis points really stop you?" he said.

The rate spike is real pressure, not a new story. What's changed is that the margin for error for smaller borrowers is now noticeably thinner, and with lenders already cutting the neocloud field from 50 to 20, the companies that can't prove contracted revenue are running out of road.