Largest AI Data Centers by Estimated Power Draw (October 2026)

xAI's second Memphis site, listed by Epoch under the owner SpaceXAI, with Anthropic, Cursor and SpaceXAI as users. Epoch estimates 946 MW of IT power across two operational buildings (1,276 MW of total facility power) and about 1.1 million H100-equivalents on Nvidia B200 and B300 chips.⁵³ SpaceX's S-1 filing says the first cluster of about 110,000 GB200 processors and about 210 MW of compute power came online in 91 days, followed by 110,000 GB300 processors and 220 MW in 64 days.⁶ To power it quickly, xAI installed natural gas turbines just across the state line in Mississippi; Epoch projects 1,531 MW by Q1 2027.⁵

Amazon's Indiana campus for Project Rainier, used by Anthropic. Epoch estimates 910 MW of IT power from 16 operational buildings and about 686,000 H100-equivalents on Amazon Trainium2 chips; like many Amazon sites, it has no external cooling plant and relies on direct air-based cooling, which makes it harder to read from above.⁷⁴ Amazon describes Project Rainier as a cluster of nearly half a million Trainium2 chips and names St. Joseph County, Indiana, as one of its sites.⁸ Cleared land for 14 more buildings is why Epoch projects 1,925 MW by Q1 2028.³

Built by QTS with Microsoft as the tenant in Fayetteville, Georgia, and likely used by OpenAI and Microsoft, per Epoch, which estimates 636 MW of IT power and about 769,000 H100-equivalents on Nvidia B200 chips.⁹ Epoch dates buildings 1 and 2 to October 2025 and buildings 3 and 4 to June 2026.³ Microsoft says the site began operating in October as the second Fairwater datacenter, with NVIDIA GB200 NVL72 racks, a two-story layout and liquid cooling, linked to its other Fairwater sites over a dedicated AI wide-area network.¹⁰ Epoch's timeline has no future-dated expansion for it.

The highest-ranked site outside the United States, in Johor Bahru, Malaysia. Epoch doesn't know who owns the AI chips; it estimates 473 MW of IT power and about 584,000 H100-equivalents on Nvidia H100 and B300 chips, and leaves out the campus's three Phase 1 buildings because Phase 2 was reported as the first in the region built for large-scale AI computing.¹¹ Its latest timeline entry (Oct 2, 2026) counts ten buildings as operational, an estimate based on the state of buildings 7–10 in April 2026 plus time to finish and ramp up.³

Meta's cluster in New Albany, Ohio, which Epoch calls unusual for its ad hoc build: a patchwork of weatherproof tents, colocation facilities and conventional buildings along the business park. Epoch estimates 471 MW of IT power and about 536,000 H100-equivalents on Nvidia B300 chips.¹² Its Aug 31, 2026 entry counts four tent buildings as running, from infrared drone video that also suggested about 138.8 MW of the on-site Socrates South gas plant's 200 MW nameplate was operating.³ Epoch projects 1,022 MW of IT power by Q3 2028.¹²

A second DayOne campus in Johor, Malaysia, again with no chip owner listed by Epoch, which estimates 422 MW of IT power and about 481,000 H100-equivalents on Nvidia B300 chips.¹³ The figure rests on Epoch's assumption that buildings 7 and 8 were operational by Oct 1, 2026: they were roofed by March 2026 and their cooling and backup power looked complete in August, so Epoch allowed two months of commissioning. Its first estimate for the site was cross-checked against DayOne's report of 390 MW "in service" across its two Johor campuses in January 2026.³

The first Stargate site, in Abilene, Texas, with chips owned by Oracle and used by OpenAI. Epoch estimates 421 MW of IT power from four operational buildings and about 509,000 H100-equivalents on Nvidia B200 and B300 chips.¹⁴ The same timeline entry gives 590 MW of total facility power for those four buildings: the same estimate expressed two ways, IT load versus the whole facility including cooling, a ratio of 1.4, Epoch's default peak PUE outside the hyperscalers.³⁴ Developer Crusoe says construction began in June 2024 and the planned campus has eight buildings.¹⁵ Epoch projects buildings 5–8 online around Nov 1, 2026, for 843 MW IT (1,180 MW facility).³

Microsoft's site in Mount Pleasant, Wisconsin, likely used by OpenAI and Microsoft. Epoch estimates 369 MW of IT power from one operational building and about 446,000 H100-equivalents on Nvidia B200 chips, dating Building 1 to April 16, 2026 from a Microsoft post.¹⁶³ Microsoft describes 315 acres and three buildings with 1.2 million square feet under roof, cooled by a closed-loop liquid system.¹⁷ It carries the largest projection in this table: a revised nine-building plan behind Epoch's 2,263 MW by Q2 2028, though Epoch's methodology notes that an early run of its cooling model gave an unrealistic 1.5 GW here, having mistaken air-cooled condensers for cooling towers.³⁴

Google's site in Pryor, Oklahoma, with Google DeepMind as a "speculative" user in Epoch's tags. Epoch estimates 368 MW of IT power and about 637,000 H100-equivalents on Google TPU v5p and v5e chips, more compute than several larger sites because the compute figure depends on the chips Epoch assumes.¹⁸ Its estimate jumped from 53 MW to 368 MW when Building 2, retrofitted from a warehouse, became operational in September 2026, confirmed by drone imagery of its cooling-tower fans spinning.³

xAI's first Memphis cluster, built in a former Electrolux building, per Epoch, which estimates 340 MW of IT power and about 276,000 H100-equivalents on a mix of Nvidia H100, H200 and B200 chips.¹⁹³ SpaceX's S-1 says the first cluster of about 100,000 H100 processors and about 130 MW of compute power came online in 122 days.⁶ Anthropic said in May 2026 that it had agreed to use all of the site's compute capacity, "more than 300 megawatts" and over 220,000 NVIDIA GPUs.²⁰³ Epoch projects no further growth.¹⁹