• Data center (bar). A building or campus-level AI facility.
• IT power (bar length). The electrical power capacity needed for a facility's computing hardware, in megawatts. The chart ranks facilities by this metric.
• Current and projected IT power. "Current IT Power" is the power a facility can draw today; "Projected IT Power" is the power expected once the facility's planned buildout is complete. Both appear when you hover over a bar.

• These figures are estimates, not disclosed specifications. Epoch reports in its data-centers methodology that about 80% of the time an IT power estimate falls within a factor of 1.4 of the true value, so individual numbers should be read as approximate.
• The chart shows power capacity, not actual power draw. A facility does not necessarily run at full capacity at all times.
• Coverage is deliberately scoped: Epoch focuses on facilities that became operational in 2024 or later (with some exceptions), and includes planned data centers only when construction has started and the project has a significant chance of being completed. The threshold for inclusion is roughly 100 MW of IT power at full buildout. Most tracked facilities are in the United States; to keep coverage globally balanced, Epoch uses a lower threshold for facilities outside the US.
• Even within that scope, coverage is not comprehensive. Smaller data centers, especially those consuming less than 100 MW, are harder to find; together, these can add up to a significant amount of power consumption.
• A whole data center is not necessarily devoted to a single AI training run. Facilities can run many workloads in parallel or be partitioned across users.

Epoch identifies candidate data centers from company announcements, news outlets, third-party databases, social media, and our GPU Clusters database (for example, OpenAI's Stargate site in Abilene was already tracked there). We then research each site using commercial and public satellite imagery, permitting documents, company disclosures, and, in some cases, drone or thermal imagery.

IT power is estimated from the best available evidence. Sometimes that means inferring it from the facility's visible cooling equipment. When the type and number of chips are known, we instead multiply the chip count by the power each chip's server draws. The second method of estimation is more reliable, but harder to obtain. The cooling system's own power draw is not itself counted in IT power. Total facility power is a separate, derived figure: IT power multiplied by a peak power-usage effectiveness (peak PUE), a unitless ratio, for which we default to about 1.2 for hyperscalers such as Google and Amazon and 1.4 otherwise, unless facility-specific evidence points elsewhere.

Read the full methodology