What Is the Biggest AI Data Center in the World? Check the Meter

EXPLAINED · Mechanism · Last reviewed: 28 Oct 2026 · Next review: after Epoch AI’s next site update and Louisiana’s ruling on Entergy’s power plan for Meta’s Hyperion

Measured by power that is actually switched on, the biggest AI data center in the world in October 2026 is Amazon’s campus in New Carlisle, Indiana, built for Anthropic as Project Rainier, at an estimated 1,037 MW of IT power, just ahead of xAI’s Colossus 2 in Memphis at about 946 MW. Measured by plans, the biggest is Meta’s Hyperion in Richland Parish, Louisiana, which Meta says will reach 5 GW of compute capacity, but none of it is running yet.

Why an investor cares: the size in the headline is rarely the size that earns money. A campus bills when its power is switched on, so the gap between announced and running megawatts is time the owner pays for and others get paid in. This page builds on what an AI data center is and asks one question: which AI data center is actually the biggest, and by what meter?

What is the biggest AI data center in the world?

By the measure that matters most, power that is running today, it is Amazon’s campus in New Carlisle, Indiana. Amazon calls it Project Rainier and built it for Anthropic. Epoch AI, a research group that tracks the largest AI sites using satellite images, permits and company disclosures, estimates it at about 1,037 MW of IT power in October 2026. That is the power its computers can draw, before cooling. Add cooling and other overhead and Epoch puts the whole facility at about 1,244 MW.

For a sense of scale, Epoch notes that the city of San Diego draws about 800 MW on average. One AI campus in rural Indiana now has the electrical capacity of a large city.

Second is xAI’s Colossus 2 in Memphis at about 946 MW, then Microsoft’s Fairwater Atlanta at about 636 MW. Fourth is a campus most people outside Asia have never heard of: DayOne’s Nusajaya site in Johor, Malaysia, at about 473 MW. The table shows the rest of the top group, with how much of each planned build-out is running (our arithmetic).

IT power is the power the computers can draw, before cooling and overhead. Estimates from Epoch AI’s dataset updated 26 October 2026, which Epoch says land within a factor of 1.4 of the true value about 80% of the time. Meta’s 5 GW is its own plan.

Two cautions. Epoch’s figures are estimates; it says they land within a factor of 1.4 of the true value about 80% of the time. And its list is not complete. It tracks 95 sites with 14.2 GW of IT power and estimates that this covers about 43% of the world’s deployed AI compute. The five biggest sites on its list hold about 25% of the power it tracks (our arithmetic).

Why do rankings of the largest data centers disagree?

Because each picks a different meter. There are four in common use, and each crowns a different winner.

• Power switched on. New Carlisle leads. This is the only meter that maps to rent or revenue today.
• Power planned. Meta’s Hyperion leads at 5 GW of planned compute capacity. Meta announced that on 13 July 2026, alongside more than $50 billion of investment in the region. On Epoch’s own projections, which only count what it can see being built, Microsoft’s Fairwater Wisconsin is the largest planned site at about 2,263 MW by mid-2028.
• Chips and compute. Colossus 2 leads at about 1.11 million H100-equivalents, against about 818,000 at New Carlisle, even though it draws less power. Its Nvidia B200 and B300 chips deliver more compute per megawatt: about 1,175 H100-equivalents per MW, against about 790 at New Carlisle (our arithmetic). See what a GPU cluster is.
• Land and floor space. Microsoft said in September 2025 that Fairwater Wisconsin covers 315 acres with three buildings and 1.2 million square feet under roof, and called it the world’s most powerful AI datacenter. Epoch describes Hyperion’s tract as more than three times the size of New York’s Central Park.

Company figures and outside estimates can also differ for the same site. SpaceX’s S-1 filing of 20 May 2026 says Colossus II brought online a first cluster of about 110,000 Nvidia GB200 chips and about 210 MW of compute power in 91 days, then a second cluster of 110,000 GB300 chips and 220 MW in 64 days. That is about 430 MW of what the filing calls compute power (our arithmetic). Epoch’s 946 MW is higher because it assumes the next phase, which the filing puts at over 400 MW more, is already running, and because Epoch counts all IT load. Same campus, two honest numbers, two meters.

Our view: when you read “world’s largest data center”, ask which meter and on what date. A press release usually quotes the plan. A satellite estimate quotes what is built. Only one of them sends invoices.

How much of the biggest AI data centers is actually running?

Less than the headlines suggest. Taking Epoch’s current estimate as a share of its own projected build-out for each site, Colossus 2 is about 62% switched on, New Carlisle about 54%, Stargate Abilene about 50% and Meta’s Prometheus in Ohio about 46% (our arithmetic). Fairwater Wisconsin is about 16%. Hyperion is at zero: Epoch’s satellite images from April 2026 show construction, not compute.

Against Meta’s own 5 GW plan, Hyperion’s gap is larger still. Epoch projects about 1,632 MW of IT power there by early 2028, about a third of the plan (our arithmetic). The rest is years of building and, above all, power. Our page on how long an AI data center takes to build explains why the building is the quick part.

What is the stuck step for the biggest campuses?

Getting a gigawatt to one piece of land. No grid was built to deliver that much power to a single customer, so the biggest campuses either wait for new plants and lines or bring their own.

Hyperion shows the scale. To serve Meta’s expansion, Entergy Louisiana announced on 27 March 2026 a plan for seven new combined-cycle gas plants totaling more than 5,200 MW, about 240 miles of new 500 kV transmission lines, batteries at three sites and nuclear uprates, paid for by Meta. It is the first project filed under the Louisiana Public Service Commission’s new fast-track Lightning Amendment. Meta has also committed to help fund up to 2.5 GW of clean and renewable power. Our page on how hyperscalers buy power explains these deals.

Colossus took the other route. SpaceX’s S-1 describes Colossus II as data centers in Memphis, Tennessee, and Southaven, Mississippi, and Epoch’s site notes record gas turbines across the state line being connected as the first cluster came online. Speed came from controlling the power, not just the building.

The pattern repeats in Asia. Malaysia’s utility TNB told analysts on 28 August 2026 that it had 61 secured data center projects with 8.35 GW of maximum demand, but actual data center load in June 2026 was 1,256 MW. Our Malaysia Pulse on Johor’s grid connections covers that wait in detail.

Who gets paid when a data center gets this big?

• Utilities and power-plant builders, first. Entergy’s plan for Hyperion is paid for by Meta, and Entergy puts the combined customer benefit of its two Meta agreements at about $2.65 billion. The wires and turbines get built and paid for before a single rack bills.
• Chip and equipment sellers. Amazon said in October 2025 that Project Rainier went live with nearly half a million of its own Trainium2 chips. Colossus II’s first two clusters alone used about 220,000 Nvidia chips (our arithmetic, from the S-1). See who buys AI GPUs.
• Builders and financiers of the shell. In October 2025 Meta moved Hyperion into a joint venture 80% owned by funds managed by Blue Owl Capital, to fund about $27 billion of development costs; Blue Owl put in about $7 billion of cash and Meta took a one-time distribution of about $3 billion. Our data center financing page and our Wire on Hyperion’s bond follow that money.
• Landlords and developers. Not every giant campus is self-built. Epoch’s notes on Stargate Abilene cite announcements from its developer, Crusoe, while Oracle owns the chips and OpenAI uses them, and the two Johor campuses carry the name of their operator, DayOne. See who builds AI data centers.
• Last: the owner of the compute. It earns nothing on the unfinished share of a campus. At Hyperion that is all of it for now, which is why the size of the plan and the size of the bill can drift apart. Our data center cost page shows where the money goes.

What is the biggest AI data center in Asia?

Johor in Malaysia holds Asia’s biggest AI campuses on Epoch’s list. DayOne’s Nusajaya site, at about 473 MW of IT power, ranks fourth worldwide, and its Kempas site, at about 422 MW, ranks sixth. Together they are about 895 MW (our arithmetic). Epoch counts about 1,138 MW across all Malaysian sites it tracks, about 8% of its global total, second only to the United States at about 84% (our arithmetic). Epoch notes that it uses a lower bar outside the US, so read those shares as a floor for tracked sites, not a census. Our page on where AI data centers are located explains why Johor became the hub.

TNB’s own numbers point the same way: of Malaysia’s 8.35 GW of secured data center demand, 5.6 GW, or 67%, is in Johor. TNB also showed a Cushman & Wakefield ranking that puts Johor first among Asia-Pacific markets that operators prefer, ahead of Sydney and Shanghai.

Singapore is the opposite case. It rations new capacity, and in August 2026 its Economic Development Board and IMDA provisionally allocated 200 MW in total, 50 MW each to four operators. That entire national round is about a fifth of the power Epoch estimates is running at New Carlisle alone (our arithmetic). See our Singapore Pulse on its power plan for data centres and what Keppel DC REIT does.

Australia has no campus on this scale yet, and Epoch tracks only about 41 MW of AI sites there. The plans are large, though: NEXTDC said in its FY26 results that it has bought a site for M5 Melbourne and begun planning for up to 1.2 GW of IT capacity. See what NEXTDC does.

What would change this ranking?

• Hyperion switches on. If Entergy’s plants and lines arrive on schedule, Meta’s campus could pass every site on this list. Watch Louisiana’s ruling on the plan.
• Colossus 2’s next phase lands. The S-1’s extra 400 MW or more would make it the largest running site, if New Carlisle does not grow faster.
• Better data. Epoch covers about 43% of deployed AI compute by its own estimate. Sites in China or elsewhere with less satellite and permit data could rank higher than shown.
• Distributed builds. On Epoch’s list, Google’s AI capacity sits across several sites of about 300 to 370 MW each, in Oklahoma and Ohio, rather than one giant campus, so a per-site ranking understates it.

How to read a “world’s largest data center” claim

• Which meter? Running power, planned power, chips or floor space are four different answers.
• IT or facility power? Facility power includes cooling and overhead and can be 20–40% higher than IT power on Epoch’s default assumptions. See PUE.
• What date? A gigawatt by 2030 is a plan; a gigawatt today is a campus.
• Who is the source? Company plans, regulatory filings and satellite estimates each have a different bias.
• Where is the power coming from? A named utility, plant and date is the strongest sign the number will become real.

FAQ

Related: what is an AI data center; hyperscale vs colocation; AI training; who the hyperscalers are; Meta’s unstarted data center leases; what Applied Digital does.

Stack doors

Layer: AI Data Centers. Also: Hyperscalers, Neoclouds, Semiconductors. Markets: US, Malaysia, Singapore, Australia.

Get the Wire — short daily notes when a giant AI campus gets its power, its chips or its money.

References