AI data centers: a census of the campuses

Series 2 · AI campus pipeline · records to 16 Oct 2026

Where capacity for AI training and inference is running, being built or has been announced — 634 campuses, each with its operator, country, status and the capacity its source publishes, counted in megawatts and summed only where a number exists.

§2Where the capacity is

The United States dominates: 233 campuses and 50.6 GW of published capacity, 57% of the total. Texas alone accounts for 43 campuses and 14.3 GW, more than any country outside the US. Several of the largest sites sit far from the established data center hubs — Richland Parish in Louisiana, Laramie County in Wyoming, Shackelford County in Texas — which is where campuses of this size found land and grid capacity.

Outside the US, Australia (7.0 GW), India (6.4 GW) and Indonesia (4.7 GW) lead. Taken together, the South, East and Southeast Asian markets in the census add 21.3 GW across 211 campuses. Europe totals 5.9 GW across 98 campuses: many sites, but the median European campus with a published figure is 41 MW, tens of megawatts rather than gigawatts.

§3Built, building, announced

Most of the published capacity is not running yet. Campuses under construction add up to 48.3 GW and announced ones to 31.6 GW, against 9.1 GW for operational sites carrying the AI flag. Part of that gap is definitional — operational facilities enter only on the AI flag, while projects in development also enter on size — but it also reflects how recent the build-out is.

Announced capacity deserves the most caution. An announcement is a plan with a number attached; financing, permits and above all grid connections decide whether it becomes a building. The capacity statements below show both directions: regulators approving supply for one campus, a tenant warning of delay at another.

§4The largest campuses

Meta Hyperion (Richland Parish) leads the census at 5,000 MW (construction). OpenAI Project Camellia (Effingham County) follows at 3,210 MW; the Georgia Public Service Commission's approval of 3,200 MW supply for an OpenAI project in Effingham County, listed in the statements below, appears to concern the same site. Of the twenty largest, 15 are in the United States; the rest are in Australia, India and United Kingdom.

Ownership is mixed. Hyperscalers building for themselves (Meta, Google) sit beside developers that lease to AI tenants, companies that started in bitcoin mining (IREN, Galaxy, Hut 8) and GPU clouds building their own campuses (Crusoe, Nebius). Where a source names the tenant, it appears in the operator column.

§5Public capacity statements

Alongside the facility records, the census keeps a short register of public statements about capacity: approvals, power deals, financings, allocations and delays, each quoted with its own MW figure and source. These are kept apart from the totals because several describe campuses that are already rows above.

The largest single figure is the 8,000 MW single-tenant campus in Pike County, Ohio announced by Nvidia / SB Energy / OpenAI. In the other direction, Oracle warned it may defer rent on the 2.45 GW Project Jupiter campus in Doña Ana County, New Mexico if a stalled gas pipeline delays power. Singapore allocated 200 MW of power on Jurong Island: DC-CFA2, 50 MW each to four operators.

§6The full census

Every campus in the census, sorted by published capacity. Search by name, operator or city, or narrow the list by status, country or region, capacity class and the rule that brought the campus in; the census field above highlights the same selection. Machine-readable copies: JSON (with every source) and CSV. For every data center in the catalogue — not only the AI-oriented ones — the megawattlas.com atlas maps them.