The Largest AI Data Center Doubles Its Power Every Ten Months

Epoch AI has been tracking one narrow but telling number: the power capacity of the single largest AI data center in operation. Since mid-2024 that record has doubled about every ten months. In August 2024 the leader was Google's Papillion site at 237 MW. By February 2025 it was xAI's Colossus 1 at 278 MW, the Anthropic and Amazon build at New Carlisle reached 398 MW that June, and a year later Colossus 2 crossed roughly 946 MW. On this curve Epoch expects the frontier to pass 1,500 MW by early 2027.

The reason to watch power rather than chip counts is that power is the hard constraint. A gigawatt-scale campus needs new transmission lines, substations, and often its own generation, and none of that appears on the timeline a training run wants. Epoch's data insight frames the doubling as a supply-side race that is now measured against the grid, not against the fab. A single site drawing close to a gigawatt uses about as much electricity as a mid-size city, so the growth is starting to compete directly with everything else connected to the same network.

Why it matters

If you plan capacity for a utility, negotiate interconnection, or forecast AI training budgets, this sets a concrete pace to plan against: the next record site will want more than a gigawatt within roughly a year. The bottleneck is shifting from buying accelerators to securing power and the grid connections to deliver it.