Show the site for
AI Policy HandbookLearn
Start hereHow AI worksWhich AI?Best AI for artworkPromptingChat privacyScamsMyths vs factsGlossaryAI explainedAI historyFeatured tool: GitHubTry AI
Try itPrompt libraryPrompt contestJob helper (beta)AI MusicNews & data
NewsHow AI helped meArchiveIncidentsSafety testsChartsHow often AI is rightData centersAI by countryJobs & economyCompanies & modelsAI systemsBuild & rules
Build with AIAI safetyRules & jobsAI Policy HandbookIs your site AI-ready?
Data centers · Feature
Computers started on liquid, moved to air, and are now going back to liquid.[1,2]
Oh My AI · Facts as of the dates shown, checked Oct 25, 2026. Numbers in brackets link to the sources list.
Photo: Oak Ridge National Laboratory, CC BY 2.0, via Wikimedia CommonsIBM’s big computers of the 1960s to early 1990s were cooled partly with water.[1,3] In the mid-1990s, lower-power CMOS chips let IBM’s servers switch to air.[1] Now the newest AI racks are designed to be cooled entirely by liquid, with warm-water designs, sealed closed loops, and in some places waste heat sent to homes nearby.[2,4,5]
“By 2023, hyperscale and colocation account for 84% of the 66-billion-liter total, while internal data centers fell to just 12%, driven by water efficiency improvements.”[6]Lawrence Berkeley National Laboratory, on U.S. data centers’ direct (on-site) water use, which it says is “primarily associated with cooling infrastructure.” As of Dec 2024.
Timeline
IBM’s System/360 Model 91 used a hybrid air-to-water design: a coolant distribution unit with pumps and a heat exchanger moved 40–79 kW of its heat into building water.[1]
Front panel of an IBM System/360 Model 91, as displayed at the Living Computer Museum, Seattle. Photo: MBlairMartin, CC BY-SA 4.0, via Wikimedia Commons
IBM’s Thermal Conduction Module, used in its 3081 and later ES/9000 systems, carried chip heat through metal into water-cooled plates.[1]
Diagram drawn by us from the source cited.
The Cray-2 supercomputer’s circuit boards were immersed in Fluorinert, a liquid that does not conduct electricity.[7]
A Cray-2 supercomputer at the Museum für Verkehr und Technik, Berlin, 2022. Photo: Fridolin freudenfett, CC BY-SA 4.0, via Wikimedia CommonsIn IBM’s water-cooled ES/9000 processors, about 80% of the heat went to a closed-loop water system, supplied at 21.7°C (about 71°F). As of July 1992.[3]
Diagram drawn by us from the source cited.
Lower-power CMOS chips let IBM’s servers go fully air-cooled.[1]
Under the raised floor of a data center, 2008. Photo: Adamantios, CC BY-SA 3.0, via Wikimedia Commons
Meta’s Luleå, Sweden data center went live in June 2013 using “chilly Nordic air” to cool its servers and 100% renewable local hydroelectric power.[8]
In early tests, Meta said it averaged a PUE “in the region of 1.07” (as of Jun 12, 2013).[8]
Meta’s data center in Luleå, Sweden, photographed in 2016. Meta says it is cooled with outside air and runs on local hydroelectric power.[8] Photo: Christopher Down, CC BY 4.0, via Wikimedia Commons
NREL’s data center in Golden, Colorado uses warm-water liquid cooling, and its captured heat warms the building’s offices and other spaces.[9]
NREL reports an annualized PUE of 1.036, compared with an average of around 1.8 that it cites for data centers (undated page, checked Oct 25, 2026).[10]
Unveiling the warm-water, liquid-cooled Peregrine supercomputer at NREL’s data center in Golden, Colorado, 2013. Photo: U.S. Department of Energy, Public domain, via Wikimedia Commons
Whole servers can sit in a liquid that does not conduct electricity.[11]
Microsoft reported in April 2021 that two-phase immersion cut an individual server’s power use by 5% to 15% in tests in Quincy, Washington.[11]
An immersion-cooling tank, 2017. Servers sit in a liquid that does not conduct electricity.[11] Photo: Rolf Brink, CC BY-SA 4.0, via Wikimedia CommonsORNL’s Frontier supercomputer is cooled with warm water, targeted at about 90°F (32°C), so it needs no refrigeration and is cooled by evaporative towers instead (as of Sep 10, 2024).[12]
Frontier at Oak Ridge National Laboratory, 2022. It is cooled with warm water and no chillers.[12] Photo: Oak Ridge National Laboratory, CC BY 2.0, via Wikimedia Commons
Meta says its Odense, Denmark data center delivers about 165,000 MWh of surplus heat a year to Fjernvarme Fyn’s district heating network, enough for about 9,000 households (as of Feb 2025).[5]
Diagram drawn by us from the source cited.
Nvidia says its Vera Rubin rack is 100% liquid-cooled, with cable-free, hose-free and fanless trays, and accepts coolant as warm as 45°C (113°F). Nvidia’s own figures, as of Mar 16, 2026.[2]
Diagram drawn by us from the source cited.
Some companies are now testing computers in orbit. Google says satellites in a dawn–dusk orbit would get near-constant sunlight for power.[13]
Starcloud says its satellites shed heat by radiating it into space, with no water.[14]
Google lists challenges including thermal management, on-orbit reliability, radiation and launch costs.[13]
Starcloud-1 carried an Nvidia H100 GPU into orbit in November 2025.[15]
On Oct 1, 2026, Google said a Project Suncatcher prototype with its TPU chips, built with Planet, had reached orbit on the Transporter-18 rideshare mission.[16]
SpaceX applied to the FCC on Jan 30, 2026 for a “SpaceX Orbital Data Center system” of up to one million satellites. The FCC accepted the application for filing on Feb 4, 2026, which is not an approval.[17]
All are early tests or plans as of Oct 25, 2026.[16,15,17]
Diagram drawn by us from the sources cited.Good to know
In cold-plate systems, coolant flows in a closed loop through metal plates mounted on the chips, rather than over the electronics.[18,4]
Industry-standard quick disconnects are designed to be “drip free” and hot-pluggable, sealing both sides when unplugged.[19]
Leak sensors in compute and switch trays report to the building management system, which can trip a rack’s power breaker and close its coolant shutoff valve.[20]
Nvidia specifies PG25, about 75% water and 25% propylene glycol, for its cold-plate loops.[18]
Industry guidelines start from deionized or reverse-osmosis water and call for ongoing checks of conductivity.[21]
Immersion systems use dielectric liquids, so whole servers can sit right in them.[11]
Leaks can still happen. Operators treat them as a risk to design for and manage, not one that has gone away.
Interactive explainer
Pick a method to see how it works, in plain steps, and why it matters. Facts as of the dates shown; checked Oct 25, 2026.
Diagram drawn by us from the sources cited.
Diagram drawn by us from the sources cited.
Diagram drawn by us from the sources cited.
Microsoft says its designs since August 2024 evaporate zero water for cooling, avoiding more than 125 million liters a year per data center, with a “nominal increase” in energy use (as of Dec 9, 2024).[4]
Diagram drawn by us from the sources cited.
NREL reports an annualized PUE of 1.036 for its data center, versus around 1.8 on average (checked Oct 25, 2026).[10]
Diagram drawn by us from the sources cited.
Microsoft reported in April 2021 that two-phase immersion cut an individual server’s power use by 5% to 15% in Quincy, Washington tests.[11]
Diagram drawn by us from the sources cited.
Meta says Odense delivers about 165,000 MWh of surplus heat a year to Fjernvarme Fyn, enough for about 9,000 households (as of Feb 2025).[5]
Diagram drawn by us from the sources cited.
Nvidia says coolant that warm lets many data centers use ambient air and closed-loop dry coolers, reducing the need for compressors, and that the power saved could run up to 10% more racks.[2]
Nvidia also says servicing time is up to 18x lower and tray assembly drops from more than 1.5 hours to about 5 minutes. Nvidia’s own figures, as of Jan–Mar 2026; it says the rack ships in the second half of 2026.[24,2]
Home › Data Centers
Hub · Data centers
A sourced guide to the buildings behind AI: what is inside them, how much electricity and water they use, where the big sites are, and which popular claims hold up. Every number links to its source.
Last reviewed Oct 3, 2026
A data center is a building (or a campus of buildings) that houses the servers, storage and networking equipment that run online services, plus the power and cooling equipment that keeps them running. The IEA describes servers, storage and network equipment installed in racks, supported by cooling, uninterruptible power supplies, backup generators and other infrastructure.1
The IEA’s breakdown of where a data center’s electricity goes is below. The shares vary by facility; the IEA gives ranges, not a single split.
| Component | Share of data-center electricity (IEA) | What it does |
|---|---|---|
| Servers | About 60% | Run the software. AI servers add GPUs and other accelerators, which draw far more power per machine. |
| Storage | About 5% | Disks and flash that hold data. |
| Networking | Up to 5% | Switches and links inside the building. |
| Cooling | About 7% in efficient hyperscale sites; above 30% in less efficient enterprise sites | Removes the heat that servers produce. |
| Power backup and other | Remainder | Uninterruptible power supplies, backup generators and other building systems. |
Source for the table: IEA, Energy demand from AI.1
LBNL sorts U.S. data-center space into internal (in-house), enterprise, colocation (space rented to many customers), hyperscale (very large sites run by the biggest cloud companies), communications service provider and edge types.2 It reports that colocation and hyperscale facilities held 74% of U.S. server capacity in 2023 and projects 85% by 2028, and that the share of server energy used in internal data centers fell from over 60% in 2014 to nearly 10% in 2023.2
Power usage effectiveness (PUE) is the ratio of a facility’s total energy to the energy used by its computing equipment; lower is better (glossary). LBNL reports an average U.S. PUE of about 1.6 in 2014 and 1.4 in 2023, and a projected range of 1.15 to 1.35 by 2028.2
The IEA reports that electricity use by accelerated servers (those with GPUs and similar chips, mostly for AI) grows about 30% a year in its Base Case, against about 9% for conventional servers.1 LBNL attributes the more-than-doubling of U.S. data-center electricity use between 2017 and 2023 largely to GPU-accelerated servers.2 The IEA says a typical AI-focused data center uses as much electricity as 100,000 households, and that the largest ones now under construction will use about 20 times that.3
LBNL estimates that inference (answering requests) accounted for nearly 60% of AI-server energy in 2023 and projects that training would account for 50–53% by 2028.2 The 2028 figure is a projection.
Confirmed Measured or reported by statistics agencies and research labs. Projections are scenarios and are called out as such.
Water use has two parts. Direct use is water a facility itself consumes, mostly for cooling. Indirect use is water consumed by the power plants that supply its electricity. The USGS explains how “consumption” (water not returned to its source) differs from “withdrawal.”8
Cooling design differs by site and climate, so one site’s figure does not predict another’s. The deeper explanation is on Is data-center water use just cooling?
Reported Nvidia’s Vera Rubin NVL72 rack puts 72 Rubin GPUs and 36 Vera CPUs into 18 compute trays plus 9 NVLink switch trays, in a rack Nvidia says weighs about 4,000 lbs.11 Nvidia says it is in full production and on track to ship in the second half of 2026.11 The figures below are Nvidia’s own, as of Oct 24, 2026; we have not seen independent measurements.
Video: “The NVIDIA Rubin Platform: Six New Chips, One AI Supercomputer” (Nvidia’s official YouTube channel; opens on YouTube). It shows the tray with no cables, hoses or fans.
Warmer coolant does not change how much heat the chips produce; it changes how much extra energy is spent removing it. Total power use still depends on how many racks a site runs.
Most announced figures are plans or company statements, not measurements, so they carry the Reported label. Capacity is quoted in the unit each source uses (gigawatts of power, not square feet) and should not be added across rows. This is a sample of widely documented projects, not a complete list.
| Country | Site or program | What the source says | Label | Type |
|---|---|---|---|---|
| United States | Stargate, Abilene, Texas (OpenAI, Oracle, SoftBank) | OpenAI said in Sept 2025 that the Abilene flagship site was up and running on Oracle Cloud, and that Stargate had nearly 7 GW of planned capacity and more than $400 billion of planned investment across its sites.14 | Reported | Company announcement |
| United States | Meta, Richland Parish, Louisiana | Meta said in July 2026 that the site will expand to 5 GW of capacity, with an investment of more than $50 billion.15 | Reported | Company plan |
| United States | Microsoft, Mount Pleasant, Wisconsin | Microsoft said on June 23, 2026 that it had completed construction of its first data-center facility there.16 | Reported | Company announcement |
| United States | AWS Project Rainier, Indiana | Amazon says the cluster uses nearly half a million Trainium2 chips and is used by Anthropic; St. Joseph County, Indiana is one of its sites.10 | Reported | Company statement |
| Ireland | National total | Data centers used 22% of metered electricity in 2024.6 | Confirmed | National statistics |
| United Kingdom | AI Growth Zones | The UK Government’s one-year review of its AI Opportunities Action Plan reports five AI Growth Zones, £28.2 billion of related investment and Isambard-AI in Bristol.17 | Reported | Government report |
| European Union | AI Gigafactories | The Commission’s July 30, 2026 call covers up to seven AI gigafactories, with €10 billion of EU and national funding and at least €20 billion of private investment sought; the EU also counts 19 AI Factories.18 | Reported | Government call (plan) |
| China | Eastern Data, Western Computing | In Feb 2022 China approved eight national computing hubs and planned ten national data-center clusters, to move computing resources toward the less developed western regions.19 | Confirmed | Government program |
| India | Google AI hub, Visakhapatnam | Google announced a $15 billion investment for a gigawatt-scale AI hub on Oct 14, 2025.20 | Reported | Company announcement |
| United Arab Emirates | Stargate UAE, Abu Dhabi | OpenAI said in May 2025 that the project is a 1 GW cluster, with the first 200 MW expected to go live in 2026; partners named are G42, Oracle, NVIDIA, Cisco and SoftBank.21 We have not confirmed that any of it is live. | Reported | Company plan |
| Saudi Arabia | HUMAIN (Public Investment Fund) | The PIF launched HUMAIN on May 12, 2025 to operate across the AI value chain, including data centers.22 The release gives no capacity figure. | Reported | Government-owned company |
| Singapore | National policy | The Ministry of Digital Development and Information said on Sept 9, 2026 that a Digital Infrastructure Bill will mandate baseline sustainability standards for all data centers and that the government is reviewing whether more capacity is needed.23 | Confirmed | Parliamentary reply |
Stanford’s AI Index 2026 reports data-center counts for 2025 compiled from Cloudscene.7 These are counts of facilities, not capacity: one hyperscale campus and one small server room each count as one, so the table says little about how much computing a country has. Only the countries the source lists are shown.
| Country | Data centers (2025) |
|---|---|
| United States | 5,427 |
| Germany | 529 |
| United Kingdom | 523 |
| China | 449 |
| Canada | 337 |
| France | 322 |
| Australia | 314 |
| Netherlands | 298 |
| Russia | 251 |
| Japan | 222 |
| Brazil | 197 |
| Mexico | 173 |
| Italy | 168 |
| India | 153 |
| Poland | 144 |
“AI is already using a huge share of the world’s electricity.”
The IEA puts all data centers (AI and everything else) at about 1.5% of global electricity in 2024.3 The growth rate is high, but the share is small today.
Confirmed
“Data centers will use the same share of electricity everywhere.”
Use is uneven. The U.S. share was 4.4% in 2023 per LBNL, and Ireland’s was 22% in 2024 per its statistics office.2,6
Confirmed
“The forecasts are settled.”
They are scenarios. The IEA’s 2035 range runs from 700 to 1,700 TWh, and LBNL’s 2028 U.S. range is 325 to 580 TWh.3,2
Confirmed
“Every AI question uses a bottle of water.”
The one published per-prompt measurement we found is a company estimate of about 0.26 mL for a Gemini text prompt, which Google says is not independently verified. See the myth page for the full discussion.9
Reported
“Data centers mostly use water that is lost for good.”
It depends on the cooling design. LBNL separates direct from indirect use, and Amazon says some of its sites use no cooling water for part of the year.2,10
Reported
“The newest data centers are all the same size.”
The IEA says a typical AI-focused data center draws as much power as 100,000 households and the largest under construction about 20 times that.3
Confirmed
What we left out. We did not include capacity figures we could not find on a primary page, and we did not rank countries by total data-center power because no single source we reviewed compares like with like. See How we verify; report an error through the corrections page.
.jpg){kind=link}
.jpg){kind=link}
