AI Data Center Building Is Set to Double Even as Credit Markets Flash Warnings
The AI data center building boom is set to double next year. Research firm SemiAnalysis expects annual additions to data center IT capacity outside China to jump from 25 gigawatts this year to about 51 gigawatts next year, Yahoo Finance reported. The capacity added in that single year would roughly match everything the world built through 2024.
About three-quarters of the projected additions come from projects already operational or under construction. SemiAnalysis counts more than 24 gigawatts across 40 campuses in advanced construction slated for delivery next year.
The race is also about speed. A modular approach, in which electrical rooms and cooling systems are assembled in factories while crews prepare the site, cuts construction of a 50-megawatt AI data hall from 22 months to 14 months. SemiAnalysis values that eight-month head start at about $200 million for the owner of the chips inside such a hall. Amazon is rolling out its own modular design at scale, and Meta has put up tent-like structures to enclose buildings faster.
Not everyone buys the forecast. Brad Gerstner of Altimeter Capital called the outlook too aggressive, expecting the industry to add closer to 25 gigawatts next year, no faster than this year.
The real bottleneck is power, not buildings. Microsoft pulled out of a lease at a planned 50-megawatt data center outside London after grid officials said electricity might not arrive until the 2030s. Microsoft said the decision involved a single site and that its partnership with Nscale continues.
Behind the building boom, the credit market is sending a different signal. The cost of insuring CoreWeave’s debt against default, through five-year credit default swaps, ended Monday near 1,090 basis points, nearly triple its level a year ago, according to Bloomberg data cited by Yahoo Finance. By Yahoo Finance’s calculations, that price implies roughly a 60 percent chance of default over five years. It is a price, not a prediction. Sellers of default insurance charge more when they are scared, but the protection has become expensive.
CoreWeave rents data center capacity to AI companies and has piled up debt to build it. Its shares sank early Tuesday to their lowest level since May 2025, extending a slide of more than 40 percent since Oct. 6, before buyers stepped in. The reversal has the makings of a bear trap, a break below support that snaps back and burns traders betting on further losses. The stock needs to hold above Monday’s low of $59.25 for the pattern to count.
The warning signs no longer stop at CoreWeave. Oracle’s default insurance costs climbed to their highest since at least 2021, and credit default swaps on Meta, Amazon, Alphabet and Microsoft all sit at their highest in months. The credit market has begun pricing risk beyond the fringes of the AI trade.
There is a counterargument. BofA Global Research called last week’s rout a positioning-driven hiccup on the fringes of the AI trade, with the sharpest pain in highly leveraged stocks such as CoreWeave and Nebius rather than the AI megacaps.
The two markets are split over the same question. Few dispute that AI demand is real. The fight is over whether that demand returns as money fast enough, and in large enough amounts, to carry the debt being piled up to serve it.