Vibeleaderboard
← All Intel
Intel / article

Largest Data Center Compute

Source
epoch.ai
Author
Epoch AI publications
Date

Context

This June 11, 2026 data insight from Epoch AI, by Ben Cottier, follows a single record: the computing capacity of the largest AI data center at any moment. Capacity here is the combined peak calculation speed of the AI chips in the facility, expressed as the number of NVIDIA H100 chips that would deliver the same, and Epoch tracks it rather than power because it is more relevant to AI capabilities. Since SpaceXAI's Colossus 1 launched in August 2024, that record has grown about 3.3 times a year, a doubling every seven months, with Anthropic-Amazon New Carlisle, Microsoft Fairwater Atlanta and Meta Prometheus each holding it at different times. The most recent observed record in the data is Meta Prometheus at an estimated 760,000 H100-equivalents in late May 2026.

The trend is fitted to only 10 past observations, and each is an estimate, mostly derived from a site's power capacity, read from satellite imagery, disclosures and permits, and the efficiency of the chips it most likely uses. When Epoch randomly perturbed those estimates 1,000 times and refitted, growth ranged from 2.3x a year at the 5th percentile to 5.1x at the 95th, which it calls an illustrative range rather than a true confidence interval. Epoch also says its dataset covers only about 26% of global AI computing capacity as of March 31, 2026, but that it prioritised finding the largest sites, and that a large facility outside the US could be missing.

Looking forward, Epoch expects no data center to have meaningfully more capacity than Colossus 2, projected at about 1.1 million H100-equivalents from July 2026, until the second half of 2027, then a return to trend in late 2027 or early 2028 as QTS Cedar Rapids and Meta Hyperion are projected to be completed. After reviewing known candidates it concludes the frontier is unlikely to grow faster than the current trend through 2028. Epoch also cautions that the largest data center is not necessarily the largest training cluster, and that the largest by compute need not be the largest by power, the measure tracked in a companion insight (epoch.ai).

More from Epoch AI publications
Recommended reads
Comments

Checking sign-in…

Loading comments…