How to read the tracker
This tracker compares AI data-center capacity across operators. The default series is IT power in megawatts, summed from the Epoch AI public campus file labeled updated 12 Oct 2026. That file covers 94 sites, 14.1 GW of IT power, and 14.6 million H100-equivalents. Epoch estimates the set covers 44 percent of global deployed AI compute as of 13 Oct 2026, with a 90 percent interval from 23 percent to 81 percent. It is not an entire cloud footprint.
H100-equivalents are the Epoch modeled conversion of those campuses. Disclosed chip counts stay on their own series and are not added into the H100-equivalent line. A campus is attributed to its hardware owner, so a lab renting the cluster does not get a second row. Blank-owner shells and speculative owners stay out. Empty cells mean no comparable disclosure, not zero.
Operating power on the 12 Oct file
Summing current IT power for confident, likely, and unlabeled owners, and leaving out speculative and blank owners, leaves nine operators on the same totals as their 24 Sep points. Google is 3,179 MW, Amazon 1,743 MW, Microsoft 1,519 MW, SpaceXAI 1,303 MW, CoreWeave 679 MW, Oracle 493 MW, Huawei 341 MW, Alibaba 169 MW, and Nebius 125 MW. The largest single campuses are still Colossus 2 at 946 MW, Anthropic-Amazon New Carlisle at 910 MW, and Microsoft Fairwater Atlanta at 636 MW.
Those figures are the Epoch estimate of IT power the covered campuses can draw now. They are not facility power, not a contracted pipeline, and not a planned end state. Google Kansas City East is already inside the Google total at 105 MW. Several named projects, including CoreWeave Lancaster and Muskogee, are in the file at zero current power.
The Meta revision, and the shells left out
Meta is the operator that moved. The 12 Oct sum is 2,244 MW and 2,489,536 H100-equivalents. That is 87 MW and 99,545 H100-equivalents above the 2 Oct point, because the 5 Oct changelog restored Meta Gallatin at 87 MW after Epoch said those metrics had been set to zero by mistake. Prometheus stays at 471 MW, down from 496 MW on 24 Sep. The H100-equivalent total is still below the 24 Sep figure of 2,523,813 because Prometheus compute was cut from 600,000 to 536,000 H100-equivalents and Los Lunas was corrected back to 87 MW.
DayOne Nusajaya at 473 MW and DayOne Kempas at about 422 MW are larger than several tracked operators, and the 2 Oct changelog raised Nusajaya from 240 MW. Both have a blank owner, so they stay out of the series. QTS Eagle Mountain is tagged Meta speculative and is at zero current power, so it is out too. The Meta Aiken change on 8 Oct raised projected power from 69 MW to 129 MW. Current power on that row is still 69 MW.
What is planned is not plotted
CoreWeave power figures are a different series from the 679 MW Epoch current total. On 7 Oct 2026 the company said it will be the sole tenant of three 80 MW buildings, 240 MW, at the AdaniConneX Taloja campus in Navi Mumbai, with a first phase expected in mid-2028 and an option to add another 240 MW. The same release restates about 4.2 GW of contracted power as of the 11 Aug 2026 earnings call. Contracted power and a 2028 phase are not current IT power, and the company record already carries that plan.
Microsoft said on 12 Oct 2026 that it will scale the Azure Maia family to multi-gigawatt deployments starting in 2027, with Maia 300 production deployments beginning that year, and that Maia 200 is already serving production workloads including Copilot. A 96-accelerator Maia 300 rack is described at 167 kilowatts. That is a company target and a custom-chip roadmap, not an addition to the 1,519 MW Epoch current total. Colossus 2 remains 946 MW in the 12 Oct file, so later chip tranches described in September are still not treated as operating capacity.
