The Largest Data Centers in the World

The largest data center on Earth, China Telecom's Inner Mongolia Information Park near Hohhot, covers roughly 1 million square meters (10.7 million square feet) and draws about 150 megawatts of power, enough to run a mid-sized city (Data Center Dynamics; WorldsTopDataCenters). But size alone no longer defines the record books. The newest wave of AI-focused campuses from Microsoft, Meta, and Google is being measured in gigawatts, not megawatts, and that shift is reshaping how these facilities are built, cooled, and powered.

What actually makes a data center "the largest"

There is no single, universally agreed metric for ranking data centers, which is why different lists produce different winners. Some rankings measure raw floor space, the square footage of white space filled with server racks. Others measure IT power capacity, the megawatts or gigawatts of electricity a site is contracted to draw for computing equipment. A few measure total campus footprint, including land reserved for future expansion that has not been built yet.

These three metrics tell different stories. China Telecom's Inner Mongolia Information Park wins on raw floor space among today's completed facilities. Switch's Citadel Campus near Reno, Nevada, wins on total planned campus scale, with up to 7.2 million square feet already built out and long-term plans for as much as 17.4 million square feet and 850 megawatts across the full 2,000-acre site (Data Center Frontier; Switch). Meanwhile, Microsoft's Fairwater campus in Wisconsin and Meta's Hyperion campus in Louisiana are racing past a gigawatt of contracted power despite occupying less physical land than either of the above, because their racks are so much denser (NextBigFuture, 2025).

A hyperscale data center is a facility built and operated at a scale designed to support cloud computing or AI workloads across tens or hundreds of thousands of servers, typically owned by a single large operator like Amazon, Microsoft, Google, or Meta rather than leased to many small tenants. That distinction matters because hyperscale campuses can standardize everything, power delivery, cooling, and networking, down to a repeatable blueprint that gets copied building after building, which is part of why they have grown so large so quickly.

Microsoft's Fairwater campus in Wisconsin already draws more than 350 megawatts with over 150,000 Nvidia GB200 GPUs installed, and the company has stated plans to expand the site's multi-location capacity past 2 gigawatts, a scale that did not exist anywhere in commercial computing five years ago (NextBigFuture, 2025).

The largest data centers in the world, ranked

The table below compares leading facilities on the two metrics that matter most: documented physical scale and power capacity. Figures reflect current or near-term buildout as reported by industry trackers and, where available, the operators themselves; AI-era campuses are expanding so quickly that any published figure should be read as a snapshot rather than a permanent ranking.

Facility	Operator	Location	Scale	Power capacity
China Telecom Inner Mongolia Information Park	China Telecom	Hohhot, China	~1M m² (10.7M sq ft)	~150 MW
Switch Citadel Campus	Switch	Tahoe Reno, Nevada, USA	Up to 7.2M sq ft built; 17.4M sq ft planned	Up to 850 MW planned
Microsoft Fairwater	Microsoft	Mount Pleasant, Wisconsin, USA	Multi-building campus	350+ MW current; 2+ GW planned
Meta Hyperion	Meta	Richland Parish, Louisiana, USA	Multi-building campus	Targeting gigawatt-scale by 2027-2028
Meta Prometheus	Meta	New Albany, Ohio, USA	Single AI supercluster	~1 GW
Google AI campus (TPU-focused sites)	Google	Multiple US sites	Multi-building campus	500+ MW AI portion within 1+ GW campus

Floor space defined the record books through the 2010s; by 2026, power capacity, not square footage, is the metric that separates leading AI campuses (NextBigFuture, 2025; Data Center Frontier).

What is the biggest data center in the world?

By physical footprint, China Telecom's Inner Mongolia Information Park near Hohhot is the largest completed data center, covering about 1 million square meters. By planned total campus scale, Switch's Citadel Campus in Nevada is larger, with a long-term buildout target of up to 17.4 million square feet, though most of that capacity is not yet constructed (Data Center Dynamics; Data Center Frontier).

Why AI training is driving a new scale of construction

For most of the cloud computing era, data center growth tracked demand for storage, web hosting, and enterprise software, a steady, predictable curve. AI training broke that pattern. Training a frontier-scale large language model requires thousands of GPUs (graphics processing units, chips built for the kind of massively parallel matrix math that both AI and physics simulation depend on) running in tight coordination for weeks or months at a time, and each generation of model has demanded roughly an order of magnitude more compute than the last.

Industry trackers now count roughly 190 gigawatts of announced hyperscale data center capacity globally across nearly 800 named projects, with about 12 gigawatts already operational and 21 gigawatts under active construction as of late 2025 (Bessemer, cited in Data Center Knowledge, 2026). Total data center capacity is projected to roughly double, adding about 97 gigawatts worldwide between 2025 and 2030, with global capacity potentially reaching 200 gigawatts by the end of the decade (Goldman Sachs; JLL, 2026).

Think of the shift like the jump from single-family homes to skyscrapers. A traditional enterprise data center was built to a fairly fixed, known load, like a house designed for a specific family. An AI training campus is closer to a skyscraper built with unfinished floors and pre-installed elevator shafts, because the operator already knows demand will keep climbing and it is cheaper to overbuild the infrastructure once than to retrofit it every 18 months. That is why campuses like Fairwater and Hyperion are announced with multi-gigawatt long-term targets attached to their very first building.

Capital spending reflects the same urgency. Combined capital expenditures from the largest AI infrastructure operators are projected to exceed $600 billion in 2026 alone, most of it directed at new data center construction and the power infrastructure to support it (Data Center Knowledge, 2026). That is a sharp break from the previous decade, when hyperscale growth was funded largely out of steadily growing cloud revenue rather than a capital race tied to a single emerging workload.

Why do AI data centers need so much more power than traditional ones?

AI training clusters run thousands of GPUs continuously at near-maximum utilization for weeks at a time, unlike traditional web hosting workloads that fluctuate with user traffic and rarely sustain peak load. That constant high utilization, combined with GPUs drawing far more power per chip than general-purpose CPUs, is why AI campuses need gigawatt-scale power contracts that dwarf earlier-generation facilities.

The cooling problem no one solved with more air

Every watt of electricity a server consumes eventually becomes heat that has to leave the building, and the physics of that heat removal have become the binding constraint on how densely a data center can be packed. For decades, the industry standard was air cooling: hot-aisle and cold-aisle containment, where fans push cold air through server racks and exhaust the resulting hot air to be cooled again, similar to how a window air conditioner cycles air through a coil. That approach works well up to roughly 15 to 20 kilowatts of heat per rack.

AI server racks blow past that ceiling. Average rack power density climbed from about 16 kilowatts in 2025 to roughly 27 kilowatts in 2026, with many AI deployments now running 50 to 70 kilowatts per rack and the newest, most densely packed systems reaching up to 246 kilowatts per rack (industry rack density surveys cited in Network World, 2026). Air simply cannot carry that much heat away fast enough; moving enough air to cool a 100-kilowatt rack would require fan noise and airflow volumes that are impractical at data center scale.

The industry's answer is liquid cooling, specifically direct-to-chip cooling, where a coolant loop runs cold plates directly against the hottest chips on a server board rather than relying on ambient air to carry heat away. Direct-to-chip systems comfortably handle 50 to 150 kilowatts per rack, and purpose-built designs stretch to around 200 kilowatts, while full immersion cooling, in which entire server boards sit submerged in a non-conductive coolant, can support 150 to 225 kilowatts per rack (Introl; Flex/JetCool, 2026). Frontier, the exascale supercomputer at Oak Ridge National Laboratory, uses a similar fully liquid-cooled design because air cooling could not keep pace with its heat output either, a constraint that scales down to conventional processors too, as explained in why thermal throttling slows devices.

The efficiency payoff is significant. Facilities that switch to direct-to-chip liquid cooling commonly report Power Usage Effectiveness (PUE, a ratio comparing total facility power draw to power actually delivered to computing equipment) between 1.10 and 1.20, compared to notably higher figures at legacy air-cooled sites, and liquid cooling can cut cooling-related energy use by 30% to 60% (Schneider Electric; Network World, 2026). A lower PUE means more of every megawatt an operator pays for actually goes toward computation rather than fighting heat, which is part of why liquid cooling has gone from a niche option to close to a requirement for any campus hosting current-generation GPUs.

The water problem hiding behind the power problem

Cooling a data center does not only draw electricity, it draws water, and that side of the resource equation has drawn far less public scrutiny until recently. Many hyperscale facilities use evaporative cooling towers, where water absorbs heat from the building's cooling loop and evaporates it into the air, a process that is energy-efficient but consumes enormous volumes of fresh water that never returns to the local supply.

The documented figures are large. Google reported consuming roughly 6 billion gallons (about 23 billion liters) of water across its global data centers in 2023, and that figure rose to approximately 8.1 billion gallons in the company's 2025 Environmental Report (Google Environmental Report, 2025). At a single site, Google's Council Bluffs, Iowa campus consumed about 1 billion gallons in 2024 and peaked at 2.7 million gallons per day during the summer, an amount the company itself compared to the daily water use of a city of roughly 25,000 people (Google Environmental Report, 2025). Industry-wide, a 2024 analysis from Lawrence Berkeley National Laboratory estimated that U.S. data centers consumed about 17 billion gallons (roughly 64 billion liters) of water directly through cooling in 2023 alone (Lawrence Berkeley National Laboratory, 2024).

Microsoft has taken a different disclosure approach, reporting overall corporate water use rather than a data-center-specific breakdown, but the company has stated it has invested more than $34 million across 76 water replenishment projects globally and cut potable water use by 97% at its Quincy, Washington facility by shifting toward closed-loop liquid cooling systems that recirculate the same coolant instead of continuously evaporating fresh water (Microsoft sustainability disclosures, cited in industry reporting, 2025).

That closed-loop shift matters because it points at the same underlying tension already reshaping the industry's approach to heat: liquid cooling systems that pull heat directly off GPU chips, described above, tend to need far less makeup water than open evaporative cooling towers, but they cost more to install and retrofit. As AI campuses scale into gigawatt territory, operators increasingly face a genuine trade-off between water-efficient closed-loop systems with higher upfront capital cost and cheaper evaporative systems that draw far more from local aquifers and municipal supplies, often in regions, like the arid U.S. Southwest, that can least afford the draw. Community pushback over water allocation has already delayed or altered permitting for several proposed hyperscale sites, making water availability, alongside grid capacity, an emerging hard constraint on where the next generation of largest data centers can physically be built.

The power grid cannot always keep up

Building the servers and the cooling loops has turned out to be the easier half of the problem. Securing enough electricity, delivered reliably, has become the primary bottleneck constraining new AI data center construction. U.S. grid interconnection queues, the process by which a new large power consumer gets approved to draw electricity from the regional grid, are now delaying projects by years in some regions, and utility operators have warned of regional capacity shortfalls as early as 2026 (TechPlusTrends; Spheron Network, 2026).

The strain is not hypothetical. Regional grid operators in the United States have issued emergency demand alerts during periods of extreme heat combined with rising data center load, and one independent market monitor described conditions as being "at a crisis stage" in mid-2026 reporting (Detroit News; Common Dreams, 2026). Aging transmission infrastructure built decades before anyone anticipated gigawatt-scale single customers is being asked to absorb loads that, in some cases, rival the electricity demand of an entire mid-sized city concentrated in one campus.

Two responses are emerging. First, in March 2026, seven major AI infrastructure companies, including Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI, signed a White House-facilitated Ratepayer Protection Pledge, committing to directly fund grid infrastructure upgrades needed to support their own facilities rather than passing those costs to ordinary electricity customers (industry reporting, 2026). Second, operators are increasingly turning to on-site power generation rather than waiting on grid connections at all; one industry report projects roughly 30% of new data center energy capacity will come from on-site generation by 2026, up from close to zero a year earlier (irecruit.co industry tracking, 2026). Both responses point to the same underlying reality: the constraint on how large an AI data center can get is no longer primarily about chips or floor space. It is about whether enough electricity can physically reach the site.

How much power does a large AI data center use?

Leading AI campuses now range from roughly 350 megawatts of current draw at facilities like Microsoft's Fairwater up to planned targets exceeding 2 gigawatts, an amount of continuous electricity draw comparable to a mid-sized city. That is roughly ten to twenty times the power draw of a large traditional enterprise data center from a decade earlier.

Common misconceptions about the world's largest data centers

Assuming physical size and computing power always move together. China Telecom's Inner Mongolia facility is larger in square footage than most AI-focused U.S. campuses, but many newer, smaller-footprint AI campuses draw comparable or greater power because GPU racks pack far more compute, and far more heat, into the same floor space than the general-purpose servers that filled earlier-generation facilities.

Treating any single ranking as permanent. Because gigawatt-scale AI campuses are announced with multi-year buildout plans attached, a facility's current capacity and its eventual planned capacity can differ by a factor of five or more. A ranking based on today's operational megawatts can look very different once a campus's later construction phases come online in 2027 or 2028.

Believing "the cloud" is placeless. Cloud computing, as covered in how cloud computing works, depends entirely on physical buildings, physical power substations, and physical water or coolant loops sitting in specific real-world locations, which is exactly why grid capacity and cooling engineering, not abstract software limits, are what now cap how fast the largest facilities can grow.

Confusing data center scale with supercomputer scale. The exaflop-class supercomputers covered in the fastest supercomputers in the world, such as Frontier and El Capitan, are built for a different purpose, running standardized simulations like nuclear stockpile modeling and climate forecasting, and are measured in FLOPS rather than power capacity or floor space. Hyperscale AI data centers, by contrast, run varied commercial and research AI workloads across many separate racks, and are sized by power and floor area because there is no single standardized benchmark analogous to LINPACK for commercial AI infrastructure.

What this means

The record for "world's largest data center" used to be a simple contest of square footage, and by that measure China Telecom's Inner Mongolia Information Park still holds the title. But the more consequential number in 2026 is power capacity, because AI training has turned electricity, not land or floor space, into the binding constraint on growth. The facilities racing past a gigawatt of contracted power, cooled by direct-to-chip liquid loops rather than air, and increasingly generating some of their own electricity on-site, represent a genuinely different category of infrastructure than the hyperscale campuses of a decade ago. Watching which operator secures the next multi-gigawatt grid connection will likely say more about the future of AI capacity than watching who breaks ground on the next building.

Q: What is the largest data center in the world by physical size? China Telecom's Inner Mongolia Information Park near Hohhot, China, is the largest completed data center by floor space, covering approximately 1 million square meters (10.7 million square feet) and drawing about 150 megawatts of power (Data Center Dynamics; WorldsTopDataCenters).

Q: Why has data center power capacity become more important than floor space? AI training workloads run thousands of GPUs continuously at high utilization, and modern AI server racks can draw 50 to 246 kilowatts each, far more than traditional enterprise racks. That density means a smaller-footprint AI campus can require more total electricity than a much larger traditional data center, making power capacity the more meaningful scale metric.

Q: Why do AI data centers need liquid cooling instead of air cooling? Air cooling works well up to roughly 15 to 20 kilowatts of heat per rack, but AI GPU racks routinely exceed 50 kilowatts and can reach beyond 200 kilowatts. Direct-to-chip liquid cooling and immersion cooling can remove far more heat per rack than air, and typically improve a facility's Power Usage Effectiveness compared to legacy air-cooled designs.

Q: Why are power grids struggling to keep up with data center growth? Gigawatt-scale AI campuses can draw as much continuous electricity as a mid-sized city, and aging transmission infrastructure and multi-year grid interconnection queues were not designed for single customers of that size. Regional grid operators have issued emergency demand alerts during 2026 as AI data center load combined with extreme heat pushed some grids toward capacity limits.

Q: Are Microsoft, Meta, and Google building the largest data centers today? By power capacity, yes. Campuses including Microsoft's Fairwater, Meta's Hyperion and Prometheus, and Google's TPU-focused AI campuses are targeting or have reached gigawatt-scale power draw, a scale that exceeded even the largest traditional hyperscale facilities of the previous decade.

---

Read next:

• The Fastest Supercomputers in the World
• How Cloud Computing Works
• Why Thermal Throttling Slows Devices