LListicle

Largest AI Data Centers (2026)

By @pickwiseresearched with AIUpdated

The largest AI data centers, ranked by the power their chips draw today on Epoch AI's estimates from satellite images and permits. Amazon's Project Rainier and SpaceX's Colossus 2 in Memphis already draw about a gigawatt each.

What we track: IT power now · Status · Open to any business · checked October 18
IT power now:
The power the site's servers draw today, by Epoch AI's estimate from satellite images and permits; cooling adds about a third.
Status:
Running, or still under construction with nothing switched on yet.
Open to any business:
Yes when a source says any business can buy capacity there; No when it is built for or taken by one user.

These numbers inform the order but don’t set it: an editor ranks the picks, weighing them against what people who use them say, since stars and downloads also reflect hype and marketing.

Status
Open to any business
  1. Built for a single customer, Anthropic, which AWS says trains and runs Claude on Project Rainier's Trainium2 chips. Epoch AI estimates 910 MW of IT power, reaching 1,037 MW from October 19 as two more buildings come online. The land is the catch: WSBT reports Amazon damaged wetlands without a permit in August 2025 and has asked to pump 35 million gallons of water a day from the site.

    Operator
    Amazon Web Services (AWS)
    Location
    New Carlisle, Indiana
    Planned IT power
    1,925 MW by Mar 2028 (Epoch AI)
    Chips
    AWS Trainium2
    Main users
    Anthropic
  2. SpaceX's IPO filing says Grok 5 trains here while Anthropic rents capacity across both Colossus sites, and Epoch AI estimates 946 MW of IT power running, about 1.5 GW by early 2027. Power is the catch: the filing says it came online almost entirely on SpaceX's own generation, and Scientific American reports Mississippi regulators counted 69 temporary gas turbines at its Southaven power site by July 2026.

    Operator
    SpaceXAI (xAI, now part of SpaceX)
    Location
    Memphis, Tennessee; gas power plant in Southaven, Mississippi
    Planned IT power
    1,531 MW by Feb 2027 (Epoch AI)
    Chips
    Nvidia GB200, GB300
    Main users
    SpaceXAI (Grok 5), Anthropic, Cursor
  3. Microsoft says the site began running in October 2025 with Nvidia GB200 racks on a dedicated fiber network that lets it train one model together with Wisconsin, and Epoch AI estimates 636 MW of IT power since two more buildings came online in June 2026. Water is the catch: Atlanta News First found meters on the QTS campus missed from billing, and QTS was billed $147,000 after a neighbor's low-pressure complaint.

    Operator
    Microsoft, on a campus built by QTS
    Location
    Fayetteville, Georgia
    Planned IT power
    636 MW, reached June 2026 (Epoch AI)
    Chips
    Nvidia GB200
    Main users
    OpenAI, Microsoft AI, Copilot
  4. Epoch AI estimates 473 MW of IT power running here, and DayOne's October 2026 IPO filing puts this campus at 47.8% of its 962 MW in service. The catch is one tenant: an unnamed short-video company, reported to be ByteDance, paid 69.2% of DayOne's revenue in the first half of 2026, and the filing warns export controls could hit it. DayOne leases whole data halls to large tenants, not GPUs, through info@dayonedc.com.

    Operator
    DayOne Data Centers (Singapore)
    Location
    Nusajaya Tech Park, Johor, Malaysia
    Planned IT power
    473 MW, no further phases listed (Epoch AI)
    Chips
    Nvidia H100, B300 (Epoch AI estimate)
    Main users
    ByteDance (reported)
  5. Meta's first gigawatt AI cluster is spread across regular data center buildings and weatherproof tents in New Albany. Epoch AI estimates 471 MW running as of August 2026, under half of the 1 GW goal. Power is the catch: SynMax drone imagery in August found the northern halls finished but not drawing power, with their on-site gas plant still under construction.

    Operator
    Meta (some buildings leased from QTS and Vantage)
    Location
    New Albany, Ohio
    Planned IT power
    1,022 MW by Q3 2028 (Epoch AI)
    Chips
    Nvidia B300 (Epoch AI estimate)
    Main users
    Meta
  6. The furthest along of OpenAI's Stargate sites: Epoch AI estimates 421 MW running, rising to 843 MW when the last four buildings come online around November 2026. The catch is its power plant. Floodlight found 10 gas turbines and 62 diesel generators went in on minor permits with no public comment, and the developers now want 41 more turbines.

    Operator
    Oracle, built by Crusoe
    Location
    Abilene, Texas
    Planned IT power
    843 MW by Nov 2026 (Epoch AI)
    Chips
    Nvidia B200, B300 (Epoch AI estimate)
    Main users
    OpenAI
  7. Epoch AI projects growth from 369 MW of IT power today to 2,263 MW by mid-2028, as nine more buildings rise on former Foxconn land. Racine County Eye notes that Microsoft's claim of the world's most powerful supercomputer has not been independently verified, and We Energies plans new gas plants as Microsoft and Vantage drive 3.9 GW of new demand, Wisconsin Public Radio reports.

    Operator
    Microsoft
    Location
    Mount Pleasant, Wisconsin
    Planned IT power
    2,263 MW by May 2028 (Epoch AI)
    Chips
    Nvidia GB200
    Main users
    OpenAI, Microsoft AI, Copilot
  8. Epoch AI counts 368 MW running at Pryor since a second building, converted from a warehouse, came online in September 2026, and Google Cloud has rented TPU v4 chips here to outside users since 2022. The catch is water and coal: Oklahoma Watch reports 1,081 million gallons consumed in 2025, with power from a utility plant that burns coal and gas.

    Operator
    Google
    Location
    Pryor, Mayes County, Oklahoma
    Planned IT power
    368 MW, no further phases listed (Epoch AI)
    Chips
    Google TPU v5p, v5e (Epoch AI estimate)
    Main users
    Google; Google DeepMind is speculative (Epoch AI)
  9. Anthropic now rents all of it: its May 2026 deal with SpaceX covers more than 300 MW and over 220,000 Nvidia GPUs, and Epoch AI estimates 340 MW of IT power with no growth planned. The site started on gas turbines. Scientific American reports flyovers found 35 running while xAI sought a permit for 15, and a promised recycled-water plant was paused partly built.

    Operator
    SpaceXAI (xAI, now part of SpaceX)
    Location
    Memphis, Tennessee
    Planned IT power
    340 MW, no growth projected (Epoch AI)
    Chips
    Nvidia H100, H200, B200
    Main users
    Anthropic (all capacity since May 2026)
  10. Core Scientific turned a bitcoin mine into what Epoch AI estimates is 262 MW of running AI capacity, and CoreWeave plans to lease it to OpenAI, The Wall Street Journal reported. The catch: it is spoken for, and weather and contractor delays already pushed back OpenAI's start dates. Businesses that want CoreWeave GPUs go through its sales team at coreweave.com, not this site.

    Operator
    CoreWeave, on a campus Core Scientific owns and built
    Location
    Denton, Texas
    Planned IT power
    282 MW by May 2027 (Epoch AI)
    Chips
    Nvidia Grace Blackwell, 50,000+ GPUs (Core Scientific, Mar 2026)
    Main users
    OpenAI
  11. Huawei Cloud opened this Inner Mongolia campus in July 2025 as its first northern site for CloudMatrix384 Ascend clusters, rented to Chinese firms, and Epoch AI estimates 184 MW now, rising to 346 MW by 2028. The catch for anyone abroad: US guidance since May 2025 treats using Ascend chips anywhere as an export-control breach, and SemiAnalysis found CloudMatrix draws 4.1x the power of Nvidia's rival rack.

    Operator
    Huawei Cloud
    Location
    Horinger, Hohhot, Inner Mongolia, China
    Planned IT power
    346 MW by Feb 2028 (Epoch AI)
    Chips
    Huawei Ascend 910C (Epoch AI estimate)
    Main users
    Huawei Cloud customers in China
  12. Epoch AI projects 1,750 MW of IT power at Project Jupiter near Santa Teresa by the end of 2028, one of the largest Stargate builds. Power is the catch: the 2.45 GW Bloom fuel-cell microgrid still awaits a state air permit, its gas pipeline slipped to February 2027, and Oracle sent the developer a force majeure notice in September, ENR reports.

    Operator
    Oracle, developed by STACK and BorderPlex
    Location
    Santa Teresa, Doña Ana County, New Mexico
    Planned IT power
    1,750 MW by end of 2028 (Epoch AI)
    Main users
    OpenAI
  13. Meta says its Richland Parish campus will grow to 5 GW of compute, the largest in its fleet, for more than $50 billion. Epoch AI expects nothing running until a first phase around January 2028. The catch is power: Entergy wants seven more gas plants worth over $16 billion on a fast-tracked review, and advocates told WWNO the deal's terms are hard to see.

    Operator
    Meta, built by DPR, Turner and Mortenson
    Location
    Holly Ridge, Richland Parish, Louisiana
    Planned IT power
    1,632 MW by Jan 2028 (Epoch AI)
    Chips
    Nvidia Rubin (Epoch AI estimate)
    Main users
    Meta
  14. Vantage's 1.4 GW Frontier campus, about 20 km from the first Abilene site, is being built for Oracle to run OpenAI's work. Epoch AI expects the first building to run around August 2027, later than Vantage's second-half-2026 delivery target. Power is the catch: it relies on an on-site gas microgrid, and Bloomberg reported Oracle trucked in gas for early work.

    Operator
    Oracle, built by Vantage Data Centers
    Location
    Shackelford County, Texas
    Planned IT power
    1,400 MW by Nov 2028 (Epoch AI)
    Chips
    Nvidia Rubin (Epoch AI estimate)
    Main users
    OpenAI
  15. Epoch AI counts no running power here yet but projects 200 MW by December 2026 and 1 GW by 2028, which would make it OpenAI's first Stargate cluster outside the US. Mubadala's chief told The National the first 200 MW would be ready in the third quarter of 2026. The catch is risk: Iran's Revolutionary Guard named the campus a possible target in April 2026, and US lawmakers are questioning the chip export approvals behind it.

    Operator
    Built by G42, operated by OpenAI and Oracle
    Location
    Abu Dhabi, United Arab Emirates
    Planned IT power
    1,000 MW by Jan 2028 (Epoch AI); first 200 MW due 2026 (G42)
    Chips
    Nvidia GB300 (planned)
    Main users
    OpenAI

How to choose

To rent compute, few of these sites are open to any business. Google Cloud sells TPU time from Pryor, Huawei Cloud rents Ascend clusters from Horinger to firms in China, and DayOne leases whole data halls in Johor. The rest are built for one lab or for the owner: OpenAI, Anthropic, Meta or SpaceXAI. Each pick names the operator, the builder and the main users, which is who to approach to sell power, land, equipment or services.

AI data centers go by several names: GPU data centers, AI campuses, hyperscale AI sites and, in Microsoft's word, AI superfactories. For an AI data center map, Epoch AI plots every site it tracks. Older "largest data centers in the world" lists rank by floor area, which puts China Telecom's Hohhot park on top; for AI, power is the better measure, since chips and cooling grow with it.

What the operators' pages leave out is power and water. Colossus and Abilene run their own gas turbines, on permits that covered fewer turbines than were running or skipped public comment, and Hyperion waits on seven new Entergy gas plants. Planned dates slip: Epoch AI expects several sites to open later than their operators say, and Microsoft's "most powerful" claim for Wisconsin has not been checked by anyone else.

IT power is what the servers draw; the whole site uses about 1.3 times that once cooling is counted (Epoch AI). Ordered by IT power running today, by Epoch AI's estimate, then sites still under construction by the power planned for them.

Numbers checked October 18, 2026.