Playbooks · Everyday tasks
Epoch AI: the largest data centers, by power
Written November 13, 2026
Epoch AI updated its estimates of the largest AI data centers by IT power capacity. The top of the list, and how the estimates are built.
The list, in November 2026
Epoch AI's ranking of the largest AI data centers puts Anthropic-Amazon New Carlisle first, at about 1,037 MW of current IT power. IT power is the electrical capacity a facility's computing hardware can draw, before cooling and other overhead. Epoch's top five, as estimates of capacity already in place:
- Anthropic-Amazon New Carlisle — 1,037 MW
- Colossus 2 — 946 MW
- OpenAI Stargate Abilene — 824 MW
- Microsoft Fairwater Atlanta — 636 MW
- DayOne Nusajaya — 473 MW
These are operational figures: power the facilities can already draw, not the size of the finished projects. For scale, Epoch notes that the city of San Diego draws electricity at an average rate of about 800 MW — less than the top two entries on the list.
Ranked by the finished build instead
Sort the same dataset by planned end-state capacity and the order changes. Microsoft Fairwater Wisconsin leads at about 2,263 MW, projected for completion in Q2 2028. Behind it: Anthropic-Amazon New Carlisle at 1,925 MW (2028), OpenAI Stargate New Mexico at 1,750 MW (2028), Meta Hyperion at 1,632 MW (2028), and Colossus 2 at 1,531 MW (2027). A site can therefore be the largest running facility and only the second-largest project at the same time, which is exactly New Carlisle's position.
The gap between the two lists is the useful part. Several of the biggest projects are less than half built, so headlines quoting end-state gigawatts describe construction plans, while the current list describes capacity that exists now.
How Epoch builds the estimates
Operators rarely publish metered capacity, so Epoch estimates it. Candidate sites come from company announcements, news outlets, third-party databases, social media, and Epoch's own GPU clusters database. Each site is then researched using commercial and public satellite imagery, permitting documents, company disclosures and, in some cases, drone or thermal imagery.
The IT power figure itself comes one of two ways. Where the cooling equipment is visible, Epoch infers power from it. Where the type and number of chips are known, it multiplies the chip count by the power each chip's server draws; Epoch calls the second method more reliable but harder to obtain. The cooling system's own draw is not counted in IT power. Total facility power is a separate figure, derived by multiplying IT power by a peak power-usage effectiveness ratio — about 1.2 by default for hyperscalers such as Google and Amazon, and 1.4 otherwise, unless facility-specific evidence points elsewhere.
How to read the numbers
- They are estimates, not disclosed specifications. Epoch's methodology puts an IT power estimate within a factor of 1.4 of the true value about 80% of the time. Read 1,037 MW as "about a gigawatt", not as a meter reading.
- Capacity is not draw. The chart shows the power a facility can draw, and a facility does not necessarily run at full capacity at all times.
- Coverage is scoped on purpose. Epoch focuses on facilities that became operational in 2024 or later, with some exceptions, and includes planned sites only when construction has started and the project has a significant chance of completion. The inclusion threshold is roughly 100 MW of IT power at full buildout, with a lower threshold outside the United States to keep coverage balanced.
- Coverage is not complete even inside that scope. Smaller data centers are harder to find, and together they can add up to a significant amount of power.
- One facility is not one training run. A data center can run many workloads in parallel, or be partitioned across users.
Related playbooks
- AlphaGenome 2 reads variants per cell type
- 3,800 hours of independent speech through a brain implant
- Cheapest way into an always-on agent
- Draft a simple business plan outline for a side business
Sources
Back to: Everyday tasks playbooks hub · Also: AIDE2's self-improvement claim vs Epoch's evaluation