The 690-Billion-Dollar Pour
Six hyperscaler programs have committed $690 billion to AI datacenters, and 74 facilities broke ground this year. What the build-out looks like when you read it as an asset class instead of a headline.
The number is easy to say and hard to hold: $690 billion in cumulative committed capex across the six largest hyperscaler programs, with 74 AI-focused facilities breaking ground in 2026 across 28 US states [1]. Project Stargate alone carries a $500 billion envelope, and its first one-gigawatt campus is already rising in Abilene, Texas [1].
An asset class with a power bill
Read as infrastructure, the pour looks like railways or telecom: enormous fixed capital, decade-long depreciation, returns that depend on utilization nobody can yet forecast. GPU depreciation schedules are being negotiated in real time; electricity futures price AI demand before regulators finish defining it. Whoever owns the concrete owns the queue for the next training run.
The bear case is the fiber glut of the 2020s: capacity whose price collapses precisely because it exists. If agentic demand arrives even a year late, $690 billion of committed steel becomes the cheapest compute in history — a catastrophe for the builders and a subsidy for everyone after them. Both outcomes reorganize the industry; only one of them is priced in.