Environment & Energy
How much water do data centers use every day?
Nobody drinks data center water. It evaporates from cooling towers and is gone, and the global footprint was an estimated 4.5 trillion liters in 2025, mostly because the electricity the machines consume takes enormous amounts of water to generate. Cooling on site alone is large enough: Google disclosed about 23 billion liters in 2023, Microsoft 6.4 billion, and US data centers collectively 64 billion. A GPT-3-class model consumes about 500 ml of water per 10 to 50 medium-length responses, electricity included. AI workloads generate more heat per rack than traditional servers, so those numbers are going up.
Roughly 143K l every second.
Source: UNU-INWEH (2026); IEA Data Centres Report 2023; UC Riverside / Li et al. (2023).
Last checked: 18 October 2026
Data centre water vs. energy, today
Water and electricity are the twin costs of running the internet. Both are under pressure as AI workloads surge, and neither is getting cheaper to supply.
What this means for you
A 2023 UC Riverside study estimates that a GPT-3-class model consumes about 500 ml of water for roughly 10-50 medium-length responses, depending on when and where it runs. That figure includes the water used to generate the electricity, not just on-site cooling. None of it is visible, none of it comes back.
AI workloads generate more heat per rack than traditional servers, so the water intensity of computation is rising. Microsoft's 2023 sustainability report showed its global water consumption increased 34% year-over-year, driven primarily by AI infrastructure expansion.
Data centers draw water from local aquifers or municipal supplies, which means they compete directly with agriculture and drinking water in drought-stressed regions. Arizona and Nevada host some of the largest US facilities, both states already managing serious water scarcity. Several US states and EU regulators are now requiring operators to disclose water usage and obtain permits.
Water consumption: the key figures
UNU-INWEH (June 2026): the electricity powering the world's data centers carried a water footprint of ~4.5 trillion liters in 2025, about 6.8 times the earlier ~660 billion liter IEA-based estimate. The gap comes mostly from a broader accounting method, not a sudden surge in actual water use
Global data centers consumed ~560 billion liters of water in 2023 (IEA); scaled to 2025 with AI-driven growth, the earlier estimate reached ~660 billion liters/year, before UNU-INWEH's broader footprint methodology revised the picture upward
IEA: a typical 100 MW US hyperscale data center consumes ~2 million liters (530,000 gallons) of water per day, including direct cooling and power generation
Microsoft 2022: 6.4 billion liters, a 34% year-on-year increase driven by AI infrastructure expansion
US data centers consumed ~64 billion liters for direct cooling in 2023; projected to double or quadruple by 2028
Training GPT-3 is estimated to have evaporated 700,000 liters of freshwater on site, about 5.4 million liters including electricity-related water (UC Riverside, 2023)
Data centers vs. water: the hidden cooling crisis of the AI age
The hidden water footprint of AI
The energy cost of AI gets most of the attention; the water cost gets comparatively little. UC Riverside researchers published 2023 findings putting the water use of a GPT-3-class model at roughly 500 ml per 10-50 medium-length responses, the equivalent of a single water bottle. Training GPT-3 evaporated around 700,000 liters of fresh water on site (about 5.4 million liters including electricity), a one-time figure, but inference runs continuously. Microsoft's Azure AI workloads drove a 34% jump in the company's water use in a single year, which is concrete evidence that scale matters.
Location, location, location
Arizona and Nevada host some of the largest US data centers. Both states are managing serious water scarcity. A large hyperscale facility drawing 15 million liters a day competes directly with agriculture and municipal water supply, so where a data center gets permitted is effectively a water-allocation decision. Microsoft committed in 2020 to becoming water positive by 2030, meaning it intends to return more to watersheds than it withdraws. Its disclosed water use rose 34% in the following year.
Data centre water consumption trends: 2020-2026
Data centres rely on evaporative cooling towers that consume 1-5 litres of water per kWh of IT load. Global data centre water consumption reached an estimated 400 billion litres in 2022, with AI inference adding a further estimated 500 ml per roughly 10-50 medium-length responses (UC Riverside, GPT-3-class model), comparable to a bottle of water for a few minutes of chatbot use.
| Year | Rate (L/s) | GL/day | Context |
|---|---|---|---|
| 2020 | 8K L/s | 0.68 GL | Pre-AI era; steady growth |
| 2022 | 13K L/s | 1.10 GL | AI-driven acceleration; Microsoft +34% |
| 2025 | 21K L/s | 1.81 GL | AI inference at scale; growing demand across all hyperscalers |
| 2026 | 143K L/s | 12.33 GL | UN University study switches to a footprint-based method, counting water embedded in electricity generation, a broader scope than direct cooling withdrawal |
| 2028 (forecast) | 235K L/s | 20.27 GL | AI data center expansion continues on the UNU-INWEH growth path toward 2030 |
When data centre water use became a public issue
- 2012Large hyperscale data centers begin using evaporative cooling at massive scale; water consumption becomes a known issue
- 2020Global data center water consumption estimated at ~250B liters/year; major operators set water efficiency targets
- 2022Microsoft water use +34% in one year; Google: 20B liters; AI workloads begin water demand step-change
- 2023UC Riverside study: a GPT-3-class model uses ~500 ml water per ~10-50 medium-length responses; Google: 23B liters; US: 64B liters direct cooling
- 2024Microsoft launches zero-water cooling datacenter design; global estimate reaches ~660B liters/year
- 2026UNU-INWEH study finds the electricity powering global data centers carried a water footprint of ~4.5 trillion liters in 2025, about 6.8 times the earlier IEA-based estimate, mostly due to a broader accounting method
- 2030Microsoft commitment: water positive by 2030; projected doubling of data center water demand without efficiency innovation
Industry and academic research on data centre water
| Year | Finding | Value | Source |
|---|---|---|---|
| 2020 | US data centers: ~35 billion liters/year direct cooling water; global ~250B liters estimated | 250.0B global liters/year (2020 est.) | FDM Group |
| 2022 | Google: 20B liters; Microsoft: 6.4B liters (+34%); global estimate ~400B liters; AI workloads begin driving step-change | 400.0B global liters/year (2022 est.) | FDM Group |
| 2023 | Google: 23B liters; US total 64B liters direct cooling; UC Riverside: a GPT-3-class model consumes ~500 ml of water per ~10-50 medium-length responses | 660.0B global liters/year (2023 est.) | FDM Group |
| 2025 | FDM 2025: ~660 billion liters/year globally; Microsoft launches zero-water cooling datacenter design August 2024 | 660.0B global liters/year (2025) | FDM Group |
| 2026 | UNU-INWEH: water footprint of global data center electricity use reaches ~4.5 trillion liters for 2025, calculated from country-level water-intensity factors applied to 448 TWh of electricity. Broader methodology than prior on-site consumption estimates | 4500.0B global liters/year (electricity water footprint, 2025) | United Nations University (UNU-INWEH) |
| 2028 | Projection: US data center water demand to double or quadruple to 128-256B liters/year by 2028 as AI expands | US projection 2028 | FDM Group |
| 2030 | UNU-INWEH projection: the water footprint of data center electricity could reach ~9.3 trillion liters/year by 2030 as global electricity demand climbs to 945 TWh | 9300.0B global liters/year (projected 2030) | United Nations University (UNU-INWEH) |
12.3 billion litres per day in perspective
4.5 trillion liters would fill about 1.8 million Olympic swimming pools. According to UNU-INWEH's 2026 figures, that same volume would cover the basic annual water needs of over 600 million people in Sub-Saharan Africa.
Training a single large AI model, such as GPT-3, evaporated an estimated 700,000 liters of fresh water on site. An adult drinks roughly 58,000 liters over an 80-year life (about 2 liters a day), so that is the drinking water of more than ten lifetimes.
Frequently asked questions
- How much water do data centers use globally per year?
- A June 2026 UN University (UNU-INWEH) study puts the water footprint of global data center electricity use at about 4.5 trillion liters for 2025, using country-level water-intensity factors applied to the electricity consumed. That figure is about 6.8 times higher than earlier estimates based only on on-site cooling, because it also counts the water used to generate the electricity. For direct on-site cooling: US data centers used roughly 64 billion liters in 2023, Google disclosed about 23 billion, and Microsoft reported 6.4 billion in 2022, a 34% year-on-year increase.
- Why do data centers need so much water?
- Most large data centers use evaporative cooling towers: water absorbs heat, then evaporates and carries it away. The process works well, but the water is permanently lost, not recycled. Beyond the facility itself, the thermoelectric power plants generating the electricity also use cooling water, which is why the footprint-based figure (4.5 trillion liters) is so much higher than just measuring what drains from the tower.
- Is AI making the water problem worse?
- Yes. GPU clusters for AI generate more heat per rack than traditional servers, so they need more aggressive cooling. UC Riverside researchers estimated that training GPT-3 evaporated about 700,000 liters of fresh water on site (about 5.4 million liters including the water used to generate the electricity). Inference adds roughly 500 ml per 10-50 medium-length responses, including electricity-related water.
Cite this rate
AnythingCounter derived rate, last checked October 2026: about 12.3 billion L per day.
Method: Based on United Nations University (UNU-INWEH) 2026, shown as a daily rate.
Suggested citation: AnythingCounter, "Water consumed by data centers," October 2026, https://anythingcounter.com/datacenter-water-per-day
Primary source: United Nations University (UNU-INWEH).
How this number is calculated, and how much to trust it
A June 2026 UNU-INWEH report puts the water footprint of data centres' 2025 electricity use at approximately 4.5 trillion litres ÷ 31,536,000 seconds = approximately 142,700 litres/second, using country-level water-intensity factors (litres per kWh) applied to UNU-INWEH's own estimate of 448 TWh for that year (the IEA's April 2026 figure on the energy page is 485 TWh, about 8% higher). This is a broader, footprint-based methodology than the earlier ~660 billion litre estimate (IEA), which explains most of the roughly 6.8-times jump: it counts water embedded in generating the electricity itself, not just what evaporates from cooling towers on-site. Google's disclosed 23 billion litres (6.1 billion gallons) of data-center water consumption in 2023 and Microsoft's 6.4 billion litres remain useful as on-site withdrawal cross-checks, though they are not directly comparable to this footprint figure. AI-specific water use: a GPT-3-class model consumes roughly 500 ml of water per 10-50 medium-length responses depending on when and where it runs, including water used to generate the electricity (University of California Riverside, 2023).
This is a footprint-based method, not a meter reading: it counts water embedded in generating the electricity, not just what leaves the cooling towers. That explains the roughly 6.8-times difference from the earlier IEA-based estimate. Corporate disclosure data (Google 2024 Environmental Report, Microsoft 2023 Sustainability Report) cross-checks the on-site portion. Academic basis: Shaolei Ren's 2023 Nature Machine Intelligence paper on AI water use at UC Riverside. All figures understate the real total, since smaller and older facilities rarely disclose water data.
Sources: see below. More detail on our methodology page.
Sources
FDM Group - Environmental Impact of Digitalisation - IEA - Energy and AI (2025): total global data center water consumption ~560 billion liters in 2023; typical 100 MW US facility ~530,000 gallons/day - UNU-INWEH - Environmental Cost of AI's Energy Use: Carbon, Water and Land Footprints (DOI 10.53328/INR26RMA002): water footprint of global data center electricity reaches 4.5 trillion liters in 2025.
Worth a look... Data center energy - E-waste generated - Sand mined for chips.