THE NUMBERS, WITH RECEIPTS.
2 Data Centers forecasts on Chronopin: cited projections and data, each with its publisher, date and quote.
Estimated by Epoch AI from the rated power (TDP) of AI chips sold by NVIDIA, Google, Amazon, AMD and Huawei, spread into calendar quarters. Datacenter power is chip power times a 2.5x overhead for servers, networking, cooling and other facility use, assuming every chip runs at full rated draw. The Q4 2025 point (dated 2025-12-31) extrapolates NVIDIA, Google and AMD sales from their average growth since Q1 2024, so it may change. For scale, Epoch compares peak demand of New York State (about 31 GW, 2025 estimate), the Netherlands (19 GW) and New Zealand (7 GW). Epoch's companion analysis puts GPUs at about 41% of facility power at peak (servers 63%, all IT equipment 71%).
“Total AI data center power capacity reached approximately 30 GW in the last quarter of 2025—comparable to peak power usage in New York State, and outstripping many developed countries.”
“In a typical frontier AI data center, total server power is 1.53x GPU power alone.”
Checked 2026-11-06
Change in cumulative rated chip power from the prior quarter, from the same Epoch AI dataset. It is chip power, not facility power: multiply by Epoch's 2.5x overhead for data center capacity. The Q4 2025 value (1.73 GW) includes extrapolated sales for some manufacturers. Epoch separately estimates that total AI computing capacity has grown about 3.3x per year since 2022, a 7-month doubling time (90% CI 6 to 8 months).
“Total AI data center power capacity reached approximately 30 GW in the last quarter of 2025—comparable to peak power usage in New York State, and outstripping many developed countries.”
“In a typical frontier AI data center, total server power is 1.53x GPU power alone.”
“We estimate that global computing capacity has been growing by 3.3x per year since 2022 (90% CI: 2.7x to 4.1x).”
Checked 2026-11-06