Market jitters, borrowing costs and power-grid delays are raising questions about the pace of AI data centre construction. One estimate puts US capacity at 183 gigawatts by 2032, requiring more than $10 trillion in investment.
Investors are questioning the financial viability of neocloud data centre developers as rising borrowing costs squeeze margins from buying and renting out chips. Firmus Technologies cancelled its initial public offering, while Microsoft walked away from a UK lease with Nscale after delays in accessing the electricity grid; local opposition also led NextDC to shelve a Melbourne project.
Stijn Van Nieuwerburgh of Columbia Business School estimates the AI boom will produce 183 gigawatts of US data centre capacity by 2032, almost five times the pre-2025 figure, at a cost of more than $10 trillion. The article notes that SpaceX chief financial officer says its data centre investments pay for themselves in less than a year, while Blackstone President Jon Gray said construction constraints make completed projects more valuable.
A world with fewer, more expensive AI data centres would require chipmakers and equipment suppliers to adapt to lower demand, AI users to recalibrate technology costs, and financiers to reconsider their exposures. The article also says AI agents use seven times as many tokens as human-led chat prompts, raising questions about how Meta Platforms will cover the running costs of its Muse assistant.
