How this list is built
Operators rarely publish how many megawatts a site is drawing. They publish plans. So the running-power column below comes from Epoch AI's Frontier Data Centers dataset (updated 19 October 2026), which estimates IT power from satellite imagery of cooling equipment, permits and company disclosures. Treat those numbers as informed estimates, accurate to tens of megawatts, not as metered readings. The announced column is what the operator or credible press reported. Where the two disagree, the gap is the point.
The ranking
| # | Site | Operator | Location | Announced capacity | Est. running IT power (Oct 2026) | Status |
|---|---|---|---|---|---|---|
| 1 | Project Rainier (New Carlisle) | Amazon (AWS); used by Anthropic | New Carlisle, Indiana, US | ~2.2 GW across ~30 buildings (reported, 2025) | ~1,037 MW | Operating since Oct 2025, expanding; Trainium chips |
| 2 | Colossus 2 | xAI, now part of SpaceX | Memphis, Tennessee, US (power in Southaven, MS) | ~2 GW per Musk (reported); Epoch projects ~1.5 GW by Q1 2027 | ~946 MW | Operating, expanding; mostly on-site gas power |
| 3 | Fairwater Atlanta | Microsoft | Fayetteville, Georgia, US | Not disclosed | ~636 MW | Operating since Oct 2025; GB200 NVL72 |
| 4 | Prometheus | Meta | New Albany, Ohio, US | ~1 GW | ~471 MW | Serving training jobs while still under construction |
| 5 | Stargate Abilene | Oracle for OpenAI; developed by Crusoe | Abilene, Texas, US | ~1.2 GW, eight buildings | ~421 MW (Epoch projects ~843 MW by Q4 2026) | Partly operating; remaining buildings due late 2026 |
| 6 | Fairwater Wisconsin | Microsoft | Mount Pleasant, Wisconsin, US | Not disclosed | ~369 MW | Went live 2026, ahead of schedule per Microsoft |
| 7 | Hyperion | Meta (Blue Owl joint venture) | Richland Parish, Louisiana, US | Up to 5 GW; ~2 GW by 2030 | Not yet operating | Under construction; full build-out ~2032 |
Hyperion is listed last because it is not running yet, but by announced capacity it is the largest project on the list. Google does not appear as a single entry because it spreads AI capacity across many large sites (Epoch lists several between 230 and 370 MW in Oklahoma, Ohio, Iowa, Nebraska and Virginia) rather than concentrating it in one campus.
1. Project Rainier, New Carlisle, Indiana
Amazon built this campus for Anthropic, and it went live in October 2025 with seven buildings and close to 500,000 Trainium2 chips, according to SiliconANGLE. The full site was reported at about 30 buildings drawing up to 2.2 GW. It is the largest site in Epoch's dataset by IT power, and it is the largest cluster on this list that does not run on NVIDIA GPUs.
2. Colossus 2, Memphis, Tennessee
xAI built Colossus 1 in a former factory in 2024 and followed it with Colossus 2 nearby. Since the SpaceX–xAI merger agreement of January 2026, the sites belong to SpaceX. Epoch estimates about 946 MW running and projects about 1.5 GW by the first quarter of 2027; Elon Musk has been reported as saying the Memphis build-out will approach 2 GW. Most of the power comes from gas turbines across the state line in Southaven, Mississippi, where a permanent plant of about 1.2 GW received an air permit in March 2026. Epoch lists Anthropic and Cursor among Colossus users, which says something about how fluid "whose data center" has become.
3 and 6. Microsoft Fairwater, Atlanta and Wisconsin
Microsoft does not publish megawatts for Fairwater, but it has published a lot about the design. The Atlanta site began operating in October 2025 with GB200 NVL72 racks in two-storey buildings, cooled by a closed loop whose initial water fill equals what about 20 homes use in a year. Satya Nadella said in April 2026 that the Wisconsin site was going live ahead of schedule, with "hundreds of thousands of GB200s" in one system. The two are linked by a dedicated AI WAN so they can run as one. Epoch puts Atlanta at about 636 MW and Wisconsin at about 369 MW.
4. Meta Prometheus, New Albany, Ohio
Prometheus is Meta's roughly 1 GW cluster, and it is a good example of how these sites actually come online: GamesBeat described it in September 2026 as still under construction while serving training jobs and adding GPUs. Epoch estimates about 471 MW running.
5. Stargate Abilene, Texas
The first site of the Stargate venture between OpenAI, SoftBank, Oracle and MGX, developed by Crusoe and operated with Oracle. The plan is eight buildings and about 1.2 GW. Epoch's site profile estimates 421 MW running, with buildings five to eight structurally complete and projected to be operational around November 2026, which would take it to about 843 MW. Trackers that rely on Oracle's own delivery statements report different percentages complete, so expect the number to move. An adjacent Crusoe campus being built for Microsoft is not part of Stargate and should not be counted with it.
7. Meta Hyperion, Richland Parish, Louisiana
The largest announced AI campus anywhere. Meta confirmed in July 2026 that Hyperion is planned for up to 5 GW and raised the investment to more than $50 billion, per CNBC. About 2 GW is expected by 2030 and full build-out around 2032. Power comes from new gas plants Entergy is building for the site, approved by Louisiana regulators in 2025. The ownership is worth noticing: Hyperion is held in a joint venture with funds managed by Blue Owl, with Meta as tenant, which keeps much of the cost off Meta's balance sheet.
What the list tells a CTO
- Power sources are the story. Two of the top sites lean on on-site or dedicated gas. The grid cannot connect gigawatts on the schedule these projects want. More on that in AI data center power and water use.
- The boundary between owner and tenant is blurry. Anthropic uses Amazon's and xAI's sites; OpenAI uses Oracle's and Microsoft's. Capacity is traded like a commodity, which is good for availability and bad for assuming any provider's capacity is exclusively yours.
- Liquid cooling is standard at this scale. Every NVIDIA Blackwell rack-scale deployment on this list needs it; see liquid cooling.
- Plan with energised capacity. If a provider quotes you capacity "at" one of these sites, ask which building and which month.
Related
- AI infrastructure — the cluster hub
- Neoclouds compared — CoreWeave, Nebius, Crusoe and the rest of the capacity sellers
- Power and water use — the constraint behind every row in the table
- Nuclear microreactors — the long-dated answer to the power question
Frequently asked questions
What is the largest AI data center in the world?
By estimated running IT power in October 2026, the largest single AI sites are Amazon's New Carlisle, Indiana, campus built for Anthropic under Project Rainier (about 1,037 MW per Epoch AI) and xAI's Colossus 2 in Memphis, Tennessee (about 946 MW). By announced capacity, Meta's Hyperion in Louisiana is the largest at up to 5 GW, but it is still under construction.
What is the Stargate data center?
Stargate is the joint venture announced in January 2025 by OpenAI, SoftBank, Oracle and MGX to build AI infrastructure for OpenAI. Its first and best-known site is in Abilene, Texas: a campus of eight buildings planned at about 1.2 GW, developed by Crusoe and operated with Oracle Cloud Infrastructure. Epoch AI estimated about 421 MW running there in October 2026, rising to about 843 MW by the end of the year.
What is xAI Colossus?
Colossus is xAI's supercomputer in Memphis, Tennessee. Colossus 1 was built in 2024 in a former Electrolux factory; Colossus 2 is a second, larger site in the same area. Since xAI's merger with SpaceX in early 2026 the sites sit under SpaceX. Colossus 2 is powered largely by gas turbines across the state line in Southaven, Mississippi, where a permanent plant of about 1.2 GW was permitted in March 2026.
What is Meta Hyperion?
Hyperion is Meta's AI data center campus in Richland Parish, Louisiana. Meta confirmed in July 2026 that it is planned for up to 5 GW and raised the investment to more than $50 billion. About 2 GW is expected online by 2030, with full build-out around 2032. Entergy is building gas-fired plants to supply it, and the campus is financed through a joint venture with funds managed by Blue Owl.
What is Microsoft Fairwater?
Fairwater is Microsoft's name for its AI datacenter design and the sites built to it, starting with Mount Pleasant, Wisconsin, and Atlanta, Georgia. The sites use NVIDIA GB200 NVL72 racks, two-storey buildings and closed-loop liquid cooling, and are joined by a dedicated AI wide-area network so they can work as one system. Microsoft calls the combination an AI superfactory.
Sources: Epoch AI Frontier Data Centers dataset (updated 19 October 2026); Microsoft; Data Center Dynamics; SiliconANGLE; GamesBeat; CNBC; Meta Investor Relations; SEC EDGAR; Wikipedia (Colossus, Hyperion) for permitting and timeline details. Running-power figures are third-party estimates.