The largest AI data centers announced so far, ranked by the capacity their owners have officially stated, are OpenAI’s PORTS-Pike site in Ohio at about 8 GW-IT, Meta’s Richland Parish campus in Louisiana at 5 GW of compute capacity, and OpenAI’s Project Camellia in Georgia at 3.2 GW of power. Below them sits a group of campuses announced at 1 to 1.2 GW. Most of these are still being built: the figures describe full planned size, often years away, and they do not all measure the same thing. The ranking below uses only the companies’ own announcements, as of October 2026.

On this page

The ranking by announced capacity

Rank Site Owner and partners Announced capacity Basis of the figure Status and timing (owner’s statement)
1 PORTS-Pike, Pike County, Ohio OpenAI with SB Energy, Nvidia, US DOE About 8 GW IT load (GW-IT) First 800 MW in 2028; buildout through 2032
2 Richland Parish, Louisiana Meta 5 GW “Compute capacity” Ground broken December 2024; expansion announced July 13, 2026
3 Project Camellia, Effingham County, Georgia OpenAI; power from Georgia Power 3.2 GW Power Phased delivery 2028–2032; early development
4 Paducah Site, Kentucky NextEra Energy and Brookfield Over 1.2 GW Compute capacity Fully built in 2032; subject to definitive agreements
4 Milam County, Texas OpenAI, SoftBank, SB Energy (Stargate) 1.2 GW Data-center capacity Initial facilities from 2026
6 Saline Township, Michigan OpenAI, Oracle, Related Digital (Stargate) Over 1 GW Campus capacity Construction expected to begin early 2026
7 Sturgeon County, Alberta Meta 1 GW Data-center size Ground broken July 2026
7 Childress, Texas Crusoe and Lancium 1.0 GW Grid-connected campus Construction expected to start Q3 2026
7 Abu Dhabi, UAE OpenAI with G42, Oracle, Nvidia, Cisco, SoftBank (Stargate UAE) 1 GW Cluster 200 MW expected live in 2026

Ties share a rank. The Paducah figure is “over 1.2 GW” and Milam County “1.2 GW”, so they are listed together; the same applies to the 1 GW sites. The tracker at /ai-data-centers holds each record with its source and the date it was checked.

Why the figures are not directly comparable

Three different quantities appear in the table:

  • IT load. OpenAI gives PORTS-Pike in GW-IT, the power available to servers, storage and networking. A site’s grid connection is larger than its IT load because cooling and electrical losses come on top.
  • Power. OpenAI gives Project Camellia as 3.2 GW “of power” to be delivered by Georgia Power. The IT load inside that figure would be lower than 3.2 GW.
  • Compute capacity or plain gigawatts. Meta’s 5 GW at Richland Parish and the 1 GW figures for Sturgeon County, Michigan and Abu Dhabi are stated without saying whether they are IT load or grid power.

The figures are also targets for full build-out rather than capacity in service. PORTS-Pike reaches 8 GW-IT only by 2032 under OpenAI’s plan, and Camellia’s power arrives in phases through 2032.

Sites with large announced totals across several locations

Some announcements give one figure for a group of sites rather than for each campus, so they cannot be placed in the single-site ranking:

  • Stargate Oracle sites. On September 23, 2025, OpenAI said three new Stargate sites developed with Oracle, in Shackelford County, Texas; Doña Ana County, New Mexico; and Wisconsin, “can deliver over 5.5 gigawatts of capacity” between them.
  • Stargate program total. OpenAI said on October 30, 2025 that Stargate had over 8 GW of planned capacity across its US sites.
  • Nvidia’s Australian partners. Nvidia said on September 9, 2026 that it and local partners are working toward up to 2 GW of AI capacity in Australia by 2027, spread across several operators and sites.

Large sites described by GPUs, buildings or performance

Several well-known AI data centers are described by their owners in other terms than megawatts:

xAI Colossus. xAI’s own Colossus page says the system was built in 122 days, then “doubled in 92 days to 200k GPUs,” and describes “200,000 H100 GPUs in a single interconnected cluster” with “a roadmap to 1M GPUs.” xAI’s page gives GPU counts and build times rather than a power figure, so Colossus is not in the ranking above.

Microsoft Fairwater, Wisconsin. Microsoft described its Mount Pleasant, Wisconsin site in September 2025 as 315 acres with three buildings totaling 1.2 million square feet, using Nvidia GB200 racks of 72 GPUs, and said it will deliver “10X the performance of the world’s fastest supercomputer today.” Microsoft’s November 2025 post on the Atlanta Fairwater site gave rack-level density, about 140 kW per rack and 1,360 kW per row, but neither post gave a site total in megawatts.

OpenAI Abilene, Texas. The flagship Stargate campus runs on Oracle Cloud Infrastructure, and OpenAI wrote in July 2026 that in Abilene “our first campus is operating at scale.” Crusoe says it designed and built the campus.

Meta Prometheus, New Albany, Ohio. Meta referred to “our Prometheus supercluster in New Albany, Ohio” in its January 2026 nuclear announcement as one of the operations its new power deals will supply.

How to read “biggest”

“Biggest” can mean planned power, power in service, GPU count or floor area, and owners choose the measure that suits their announcement. On planned capacity stated by the owner, PORTS-Pike and Richland Parish lead by a wide margin. On GPU count, xAI’s Colossus is the only site in this list with a published figure. Chip specifications behind these counts are in /ai-chips, and the companies’ profiles are in /ai-companies.

Sources