Build times for gigawatt-scale data centers (epoch.ai)

Hyperscalers are proving they can bring gigawatt-scale AI data centers online far faster than many expected: among U.S. sites tracked by the Frontier Data Centers Hub that broke ground in the last three years, construction-to-1 GW ranges from about 1 to 3.6 years, with several projects achievable in two years or less and xAI projecting Colossus 2 in just 12 months. The hub expects the first GW-scale facilities to be operational in early 2026. That speed matters because a single gigawatt of facility power supports massive training clusters, so these timelines portend a rapid increase in available training capacity, shifting compute economics, accelerating model development, and intensifying local demands on power, cooling, water and permitting systems. The hub derives milestones from satellite imagery, permits and company disclosures, marking construction kickoff, a “first operational” point (defaulting to a 150-day gap after roof completion unless cooling/power aren’t present), and the campus-level 1 GW threshold. Important caveats: imagery cadence, permit variability and differing corporate definitions of “operational” make dates approximate; the hub’s cooling model can vary by as much as 2x though where ground truth exists its estimates are close, and errors are generally within ~1.5x when reference data is available. The dataset covers major U.S. projects (Amazon, OpenAI, xAI, Microsoft) and the code and methodology are publicly documented for scrutiny.