AI data centres, explained

What does an AI answer physically need?

Electricity, cooling, water and a grid connection. Follow one prompt into the building, then explore it yourself.

An answer needs electricity first, cooling second, and water in two places

Every AI answer is computed in a data centre (data center, in US spelling): a building of servers that turns electricity into answers, and into heat. Getting the heat out takes more electricity, and sometimes water. Water is used a second time, at the power plant that makes the electricity.

The cooling choice trades one against the other. Systems that use little water at the building tend to use more electricity, and electricity has its own water cost. How much that matters depends on where the building stands: its climate, its grid, and the water already scarce around it.

Where we can, we show ranges instead of single numbers, and say what each number counts. The story is global, but much of the measured detail comes from the United States because that is where the open studies are. We say so each time.

1 of 6

A data centre turns electricity into heat, then has to get rid of the heat

Servers take in electricity and do the computing. Nearly all of that electricity ends up as heat. AI data centres pack the heat tightly: the International Energy Agency (IEA) says a single server rack, about the size of a large refrigerator, could by 2027 have a peak power demand equal to 65 households, and that an AI data centre must remove heat equal to 30 natural gas boilers for every rack.

Around the servers sit three more things. Cooling equipment moves the heat out of the building. Backup power, batteries and generators, keeps it running if the grid falters. And a grid connection brings the electricity in, which in many places is the slowest part to get.

Power plantMakes the electricity.Uses water
electricity in
Data centre
ServersDo the computing and turn the electricity into heat. CoolingMoves the heat out of the building.May use water Backup powerBatteries and generators.
heat out

Where electricity, heat and water go. Water is used twice: at the power plant, and at the building if its cooling uses water.

Two ratios describe the overhead. PUE (power usage effectiveness) is the building’s total electricity divided by the electricity that reaches the servers; 1.0 would mean no overhead. WUE (water usage effectiveness) is the litres of water used at the building for each kilowatt-hour that reaches the servers. WUE can count water that evaporates (consumed) or all water taken in; each figure on this page says which.

2 of 6 · the interactive

Follow one prompt, then change how the building gets rid of its heat

We start from a measured number. Google reports that its median text prompt used 0.24 watt-hours of electricity and 0.26 millilitres of cooling water at its data centres. The energy is measured; the water is Google’s own calculation from that energy and its fleet-wide water efficiency. It is company-reported, for one month of 2025, and it counts the water used at the building only, not at the power plant. Google does not define how long a median prompt is.

Now scale up to a whole building and change the settings. Everything below is a simplification built from published ranges, not a forecast and not any one real facility. Start with the three buttons below, then change anything yourself.

From Google · company-reported

0.24 Wh of electricity0.26 mL of cooling water

for one median text prompt. Water at the power plant is not part of this figure.

A 100 MW data centre, for a year

Server load is 100 MW, all year: 876 GWh (a GWh is a million kWh).

Electricity

955 GWh

for every 1 kWh the servers use: 1.09 kWh

Water at the building

1,010 million litres

for every 1 kWh the servers use: 1.15 L

Water at the power plant

Not in Google’s figure

Pick a way of cooling to add it.

Total water

Not counted

Building plus power plant.

A wide bar is a range the sources allow. A narrow bar is a single value from one source.

What this does not show

Liquid cooling. Direct liquid or immersion cooling takes heat off the servers. The heat still has to leave the building through one of the routes above, and that route decides the water. As Alston et al. put it, “liquid cooled” does not say how heat is rejected from the facility.
Other climates and grids. Every band here is a US climate-zone analogue with US grid figures. We found no open range for tropical, Nordic or other climates, or for grids outside the US.
Large and hyperscale sites. The bands come from the authors’ midsize systems, which match the cooling types in the worked case. Their large-scale class, which includes hyperscale, has lower PUE (medians 1.12 to 1.25) and different cooling designs, so a real 100 MW site can sit below these bands. “100 MW” here is only the unit the readouts are scaled to.
Seawater, reclaimed water and closed-loop designs. No sourced range, so no switch.
Training, chips and construction. Only the electricity and cooling water of running servers.

3 of 6

Cooling trades water for electricity, and the power plant is where the trade lands

There are three basic ways to throw the heat away. Evaporate water in cooling towers: efficient, but the water that evaporates is gone. Use outside air when it is cool and dry enough, and a chiller the rest of the time: a hybrid. Or cool with air and refrigeration only, with air-cooled chillers or dry coolers: almost no water at the building, but the compressors work harder and use more electricity.

The IEA states the trade-off plainly: cooling equipment that minimises water use tends to consume more energy, and the other way round. It adds that the largest share of water use in the data centre value chain is upstream, in power generation.

One fully worked case

Alston, Love and Haider simulated a hypothetical 200 MW data centre in southeast Michigan, US, against a simulation of the grid there, and changed only the cooling. Their study is a preprint on arXiv, so it has not been peer reviewed. We derived the efficiency values from their table; the working is under Sources and method.

Water-cooled chiller

Electricity 3.08 TWh

Building 6.45 Power plant 4.48 Total 10.93
Hybrid: outside air plus chiller

Electricity 2.76 TWh

Building 3.20 Power plant 3.98 Total 7.19
Air-cooled chiller

Electricity 4.06 TWh

Building 0.21 Power plant 5.91 Total 6.12

Water in million cubic metres a year, consumption, for 200 MW of servers in southeast Michigan, US. Source: Alston, Love and Haider, arXiv 2609.25437 (CC BY 4.0), Table 2; we kept their figures and rounded the bar lengths.

Switching from water cooling to air cooling cut water at the building by 97% and total water by 44%. It also raised electricity by 32%, and with it the water at the power plant by 32%. The two do not simply swap one for the other: here the building’s saving was much larger than the power plant’s extra, so the total fell.

That result belongs to Michigan’s grid. In the study’s simulation, the power plants that respond to the extra load use about 1.45 litres of water per kWh (our arithmetic from its table; hydropower water is counted as zero). Other US estimates run from about 1.1 to about 4.5 litres per kWh, depending on the grid and on how hydropower is counted. Using the same worked-case figures, air cooling stops beating the hybrid on total water once the grid passes about 2.3 litres per kWh. Try it in the interactive: pick “Michigan (worked case)”, then move the third setting.

Rystad Energy, an energy consultancy, reaches a similar trade in its own global estimate: dry cooling saves about 2.15 litres of water per kWh of server load, and needs roughly 0.30 to 0.74 kWh more electricity for each kWh.

Why the headline water numbers disagree

You may have seen very different totals. They count different things, so they cannot be added or compared directly.

0.26 mL per prompt: Google, its own fleet. Water evaporated for cooling at the building only. Energy measured, water calculated from it by the company, May 2025.
About 300 billion litres a year: 472 large US data centres, 12 months to March 2026. Building plus power plant: about 74 billion at the buildings and 226 billion at power plants. Estimated, using one assumed efficiency for every site. Preprint, not peer reviewed. Its scenarios range from about 205 to 451 billion. If reservoir evaporation at hydropower dams is not counted, the total falls to about 202 billion.
66 billion plus about 800 billion litres: All US data centres, 2023. 66 billion at the buildings and nearly 800 billion at power plants, from regional grid averages. Estimated by Lawrence Berkeley National Laboratory.
222 billion litres: The world, 2025. Cooling water at buildings only; power plants excluded. An estimate by Rystad Energy.

Some count only the water a building evaporates; others count water drawn in, part of which is returned; others add the water power plants use. They cover different places and years, and use different methods: a company measuring itself, researchers assuming one efficiency for every site, analysts estimating a country. How a grid’s hydropower is counted can change a total by a third. We found two US totals, about 300 and about 860 billion litres, that we cannot bring into agreement without a shared method.

4 of 6

Demand is set to double, and the grid connection is often the bottleneck

The IEA reports that the electricity demand of data centres grew 17% in 2025, and that of AI-focused data centres 50%. Its central projection has data centre electricity roughly doubling by 2030, to around 3% of global electricity demand, with AI-focused data centres tripling.

2025485 TWh

2030about 950 TWh

Electricity used by data centres worldwide. IEA central projection, from Key Questions on Energy and AI (CC BY 4.0). A projection, not a measurement.

Per task, the news is better than the totals: the IEA says the energy used per AI task has fallen by at least a factor of ten a year in recent years, and that if all conventional web searches were done as simple AI text queries they would use under 4 TWh a year, under 1% of data centre use today. But video, reasoning and agent tasks can use hundreds or thousands of times more per query than simple text, and use is growing quickly.

The limit is often not electricity in general but a connection in one place. In several places the IEA says connection queues hold data centre capacity larger than today’s peak demand, and it names Texas, Virginia, Finland and the UK among the places with large pipelines. The queues include duplicate or speculative projects, so planners cannot tell how much will be built.

Water risk is concentrated too. The IEA finds nearly 40% of identified AI-dedicated capacity in areas of high or extremely high baseline water stress. And in the Harvard preprint of 472 large US sites, three of the 24 grid regions that host them account for 59% of the power-plant water attributed to those sites.

5 of 6

Places are starting to ask for the numbers

The IEA counts 134 policies and regulations on connecting data centres to power systems, across more than 34 jurisdictions that hold about 90% of global data centre capacity. Nearly 40% were announced or put in place in the last 12 months. The most common kind simply makes demand visible: how much, when and where.

Three examples, from three regions. Each asks a different question, and each is at a different stage. This is not a ranking.

European Union: report your water

Since 2024, operators of data centres with at least 500 kW of IT capacity must report each year to a European database, including total energy and total water input; the database works out each site’s PUE and WUE from them (EU Delegated Regulation 2024/1364). Data for single sites stays confidential; only totals and averages by country and for the EU are published. A second step, a mandatory A-to-G label for PUE and WUE, was adopted by the Commission on 21 September 2026; the database is to generate a label for each reporting site by 15 August 2027. The act enters into force 20 days after publication in the Official Journal, which we have not checked. In the first round of reports (2024), a Commission report shows an average reported PUE of 1.36 and an average reported WUE of 0.58 litres of water taken in per kWh. Those are operators’ own reports for a minority of the sites in scope, and the WUE leaves out sites that report no water use, so it cannot be set beside Google’s consumption figure.

Ireland: bring your own power

Ireland’s energy regulator published a decision on 12 December 2025: new data centres connecting at 10 MVA or more must provide generation or storage on site or nearby that matches their connection (after derating); those from 1 to 10 MVA must provide a generating unit sized to their connection; and all at 1 MVA or more must cover at least 80% of their yearly demand with additional renewable electricity generated in Ireland, on a six-year glide path. It is a rule about the grid, not about water.

India: ask before you approve

On 14 October 2026 NITI Aayog’s Frontier Tech Hub published a framework that lets state governments compare data centre proposals. Its water items ask states to work out peak-summer water demand and WUE at local conditions, including about 40°C, to prepare a 20 to 30 year seasonal catchment water balance, and in high-stress places to require independent source checks, certified meters and periodic reporting. For power it asks for a jointly validated ramp-up plan, and demand-response readiness above 20 MW. The framework is guidance, not an extra layer of regulation. We describe it in our own words; we have not asked permission to reproduce its pages.

6 of 6

What it means depends on where the building stands

Start with the person typing a question. For a simple text prompt the measured numbers are small: Google’s median is 0.24 Wh, and the IEA says a simple text query typically uses less electricity than running a television over the same period. Heavier uses are on a different scale, and the IEA notes there are no comprehensive statistics on how AI is used around the world.

The cost shows up elsewhere: at a substation, in a local water supply, on an electricity bill. Who carries it is a policy choice, not a law of nature. The IEA says the effect on bills is not clear-cut: where a new large load needs extra network or generation and the cost is shared across all users, it can raise bills; where it makes fuller use of capacity that already exists and pays its share, it can ease pressure.

Who is affected most: people near clusters where capacity has outrun grid, water or land, and customers in places where connection costs are spread across everyone. Who is affected less, directly: most people in places with no large cluster, who meet the effect through national demand, around 3% of world electricity by 2030 on the IEA’s central path.

What the evidence supports is narrower than either “AI is draining the world” or “it is nothing”. The numbers are real, they are rising, they depend heavily on cooling, climate and grid, and for most single buildings they are not published in a form anyone can check.

What we don’t know

Each open question is paired with the evidence that would settle it.

How much water one data centre uses, and from which source: Metered, audited reporting per site on a common boundary: drinking water, reclaimed water or seawater; consumption or withdrawal. The EU database and India’s framework are first steps, and the EU keeps site data confidential.
Whether the US totals can be reconciled: A shared boundary and grid-factor method, including how hydropower evaporation is counted, and a peer-reviewed version of the Harvard study.
Real operating efficiency against design values, outside the US: Measurements after commissioning. We found no open range for tropical, Nordic, arid or non-US grid conditions.
Whether large and hyperscale sites sit below the bands we show: Measured PUE and WUE of large sites by cooling type. The bands here are from midsize systems in a 2021 preprint’s simulation, chosen because they match the cooling types in the worked case.
How the water at a power plant is best counted: Plant-level and hour-by-hour data. Most of our grid factors are averages, and Michigan’s is a simulated marginal value; Alston et al. show that the generators that actually respond to a new load can differ from the grid average.
How cooling choices behave on the hottest days: Hourly or daily data, not yearly averages: the day water and electricity are scarcest may be the same day.
How much electricity AI use needs, and how fast efficiency outruns use: Disclosure by AI providers on a stated boundary. Google’s median prompt is one company, one month, text only, with prompt length undefined.
How much announced capacity will be built: Energisation and connection-queue records. The IEA flags duplicate and speculative projects.
Whether “zero water” designs work at scale: Operating data from the first sites, and what they cost in electricity. We did not include any such claim because we could not trace one to its operator report.
Whether places will adopt and enforce disclosure rules: Final texts and adoption records. The EU label is adopted but not yet confirmed in force, and the Indian framework is guidance.

Sources and method

Every figure on this page comes from one of these. “CC BY 4.0” sources are used under that licence, with credit; we simplified, rounded and recalculated where the page says so. This page is independent: it is not endorsed by the IEA, the authors or any company named. We did not reproduce any third-party chart, logo or image.

IEA, Key Questions on Energy and AI, 16 April 2026. Demand growth in 2025, the 485 to about 950 TWh projection, per-task energy, rack power and heat, and the household-television comparison. CC BY 4.0. iea.org/reports/key-questions-on-energy-and-ai
IEA, Powering Artificial Intelligence, 14 October 2026. 134 policies in more than 34 jurisdictions, connection queues, water stress share, the water-energy trade-off, upstream water, electricity prices. CC BY 4.0. iea.org/reports/powering-artificial-intelligence
Alston, Love and Haider, Data center cooling choices shift water impacts across the grid, arXiv:2609.25437, 21 September 2026. The southeast Michigan worked case (Table 2). Preprint, not peer reviewed. CC BY 4.0. arxiv.org/abs/2609.25437
Lei and Masanet, Climate- and Technology-Specific PUE and WUE Predictions for U.S. Data Centers using a Physics-Based Approach, Research Square preprint rs-769999, 3 August 2021. Simulated PUE and WUE by cooling configuration and US climate zone; the bands in the interactive are taken from the authors’ public results file (github.com/nuoaleon/Data-Center-Water-footprint, Simulation Results/UE.xlsx; no licence stated on the file, so cited, not reproduced). Preprint, not peer reviewed. CC BY 4.0. doi.org/10.21203/rs.3.rs-769999/v1
Elsworth et al. (Google), Measuring the environmental impact of delivering AI at Google Scale, arXiv:2508.15734, 21 August 2025. The median prompt (0.24 Wh, 0.26 mL), fleet PUE 1.09 and WUE 1.15 L/kWh. Company-reported. CC BY 4.0. arxiv.org/abs/2508.15734
Guidi and Dominici, The Hidden Water Geography of U.S. Hyperscale Data Centers in the AI Era, arXiv:2607.02531, 2026. About 300 billion litres, the site and power-plant split, the hydropower sensitivity, the 59% concentration, and the grid water factors we derived from it. Preprint, not peer reviewed; it applies one assumed WUE to every site. CC BY 4.0. arxiv.org/abs/2607.02531
Shehabi et al., 2024 United States Data Center Energy Usage Report, Lawrence Berkeley National Laboratory, December 2024. 176 TWh in 2023; 66 billion litres direct; nearly 800 billion litres indirect; 4.52 L/kWh. No Creative Commons licence stated: cited, not reproduced. escholarship.org/uc/item/32d6m0d1
Rystad Energy, data centre water consumption, 8 September 2026. 222 billion litres of direct cooling water in 2025; the dry-cooling trade. An estimate by a market-research firm; no terms stated: cited, not reproduced. rystadenergy.com
Commission Delegated Regulation (EU) 2024/1364, 14 March 2024. Reporting by data centres of 500 kW or more, and how PUE and WUE are calculated. EU law. eur-lex.europa.eu/eli/reg_del/2024/1364/oj
European Commission, delegated regulation establishing a common Union rating scheme for data centres (C(2026) 3472), adopted 21 September 2026; copy in the Council register, document ST 13444/2026. The act (Articles 1 and 3, the 15 August 2027 date) and, in the annexes, the A-to-G classes. Not yet confirmed in force. act · annexes
European Commission, Report on the energy efficiency of data centres in the EU (COM(2026) 500), 21 September 2026. Reported PUE and WUE averages from the first reporting round. Operators’ own, unaudited reports; WUE is total water input over IT energy, for sites reporting water use. build-up.ec.europa.eu (PDF)
Commission for Regulation of Utilities (Ireland), Large Energy Users Connection Policy, decision paper CRU/2025236, 12 December 2025. The 1 MVA and 10 MVA thresholds and the 80% renewable requirement. Published by the regulator; cited, not reproduced. cru.ie decision paper (PDF)
NITI Aayog Frontier Tech Hub, Data Centre Evaluation Framework to Help States Plan for Sustainable Growth, published 14 October 2026. Five pillars, 15 evaluation items; we describe it in our own words and reproduce nothing. niti.gov.in/node/2443

How the interactive works

The unit. A site with 100 MW of servers running all year uses 100 MW × 8,760 hours = 876 GWh at the servers, as in Alston et al., who assume a constant load.
Electricity = server electricity × PUE. Water at the building = server electricity × WUE. Water at the power plant = electricity used by the building × the grid factor you pick. Total = building plus power plant. Bands carry the low and high ends through the sums, so they are wide on purpose.
PUE and water bands come from the authors’ public results file for Lei and Masanet’s simulation. For each US climate zone and cooling configuration it gives the 5th and 95th percentile over 50 sampled designs. A band here is the lowest 5th to the highest 95th percentile over the zones in the climate you pick, so it is wide on purpose. We use their midsize systems: water-cooled chiller (their case 5), hybrid with an air-side economiser (case 3) and air-cooled chiller (case 7). PUE is highest in hot, moist zones and lowest in cold ones, and air-cooled chillers are the least energy-efficient, as the authors report.
Large sites. Their large-scale class (above 20,000 square feet, “inclusive of hyperscale”) has lower PUE and different cooling, and is not drawn here, so a real 100 MW site can sit below these bands.
The Michigan values are derived from Alston et al., Table 2: PUE = facility electricity ÷ 1.752 TWh of server electricity; WUE = building water ÷ 1.752 billion kWh. Water-cooled 1.76 and 3.68, hybrid 1.58 and 1.83, air-cooled 2.32 and 0.12. The authors took their PUE and WUE inputs from Berkeley Lab’s midsize ranges: the upper ends for the water-cooled and air-cooled cases, and, for the hybrid, the lower ends in outside-air mode (25% of the year) and the upper ends in chiller mode. With Michigan’s factor the interactive reproduces the paper’s table to within about 1%. The air-cooled water value, 0.12, is a little above the 0.10 top of the simulation’s band: the two come from different ranges.
Climate groups are our choice of the authors’ zones: mild = 3C, 4A, 4B, 4C; hot and humid = 1A, 2A, 3A; hot and dry = 2B, 3B; not sure = all 15 zones in the file. The authors’ zone map names a representative city for each (for example Miami, Phoenix, Seattle).
Grid water factors (litres of water consumed at power plants per kWh): 1.1 and 1.95 are our arithmetic from Guidi and Dominici (about 128 and 226 billion litres at power plants, over about 116 TWh of facility electricity; the first leaves out hydropower reservoir evaporation). 1.45 is our arithmetic from Alston et al.: water at the power plants that respond to the extra load in a simulation of Michigan’s grid, with hydropower water counted as zero. 4.52 is Berkeley Lab’s US data centre average for 2023. They are US figures, they are four estimates, not the range across US grids, and they are not transferable to another grid.
Google’s fleet uses its reported PUE of 1.09 and WUE of 1.15 L/kWh as they are; no climate or grid is paired with a worldwide company average.
The 2.3 L/kWh crossover is our arithmetic on the worked-case values: hybrid total water = 1.83 + 1.58g, air-cooled = 0.12 + 2.32g, equal at g = (1.83 − 0.12) ÷ (2.32 − 1.58) = 2.31.

What we left out, and why

Some widely repeated figures are not on this page because we could not trace them to a measurement we can cite: a global water total and a “litres a day per 100 MW” rule of thumb that are both attributed to the IEA but that we did not find in the IEA reports we read; an Indian national total that its publisher attributes to a market-research firm; and a 2028 bank projection we could find only in secondary coverage that disagrees with itself. We also left out company “zero water” and “water positive” claims we could not trace to an operator report.

AI tools helped draft and check this page. Every figure was traced to the source listed above. Found a mistake? Tell us; we correct in public. Last updated 2026-10-25.

More stories you can explore yourself, from a warm patch of the Pacific to what comes next.

All explainers What Just Happened