AI data-center boom faces risk of delayed overbuilding

AI’s projected scaling has so far tracked better than some earlier technology forecasts, which were associated with repeated misses in areas such as VR, flying cars, self-driving cars and NFTs. The comparison characterized AI as “very different,” while noting that expectations of a scaling plateau have not yet been borne out. Large investments alone do not prove there is a bubble; the key test is whether the industry finds useful products and services customers will pay for.

Data-center investment could still overshoot. High prices prompt new capacity, but it arrives with a delay, by which time users may have optimized demand or found substitutes. Every megawatt built so far continues to become more valuable, and compute available tomorrow is worth more than capacity arriving in two years—and much more than capacity arriving in five. The risk is that builders, expecting to rent out or sell capacity, respond to the same price signals and create a glut; the eventual oversupply could also bring a much steeper price decline than some investment models allow for.