Working Paper 11
Can the AI Buildout Pay for Itself? Demand Absorbability, Price Declines, and the Timing Mismatch in AI Infrastructure Investment
September 2026
Abstract
Can the planned AI datacenter buildout pay for itself? Auditing 52 published forecasts, we put the capital hurdle at about $3.55B per GW-year, map each $1T of cumulative 2026-2035 investment to about $250B per year of required end-market revenue, reject the Jevons backfire story for AI compute, and identify the central stranding risk as a timing mismatch: 30-year facilities funded against 5-year hardware on a 15%-per-year price decline curve.