> ## Content Index > Fetch the complete content index at: https://www.sotkn.com/llms.txt > Use this file to discover other available public pages before exploring further. # How Much Water Does an AI Data Center Use? - URL: https://www.sotkn.com/explained/ai-data-center-water-usage/ - Published: 2026-10-16T03:28:45.000Z - Updated: 2026-10-16T03:28:45.000Z - Description: An AI data center uses water mainly to cool servers. Microsoft’s fleet WUE was 0.27 L/kWh in FY25; Singapore’s PUB median is 2.1 m³/MWh. Who pays, and where wet cooling sticks. - Author: Sean - Tags: Mechanism, #desk-explained EXPLAINED · Mechanism · Last reviewed: 16 Oct 2026 · Next review: after the next Google, Microsoft and Equinix water or sustainability disclosures An AI data center uses water mainly to cool servers and control humidity; the usual measure is water usage effectiveness, or WUE, in litres of water per kilowatt-hour of IT energy. How much it uses depends on the cooling kit and the climate: Microsoft’s fully owned fleet averaged 0.27 L/kWh in FY25, Google’s consumptive fleet WUE was 1.15 L/kWh in 2023 and 2024, and Singapore’s PUB median for data centres is 2.1 m³/MWh. **In short.** An AI data center uses water mainly to cool servers. The published yardstick is water usage effectiveness (WUE): litres of water per kilowatt-hour of IT energy. Microsoft’s owned fleet averaged 0.27 L/kWh in FY25; Google’s consumptive fleet WUE was 1.15 L/kWh in 2023 and 2024; Singapore’s PUB median for data centres is 2.1 m³/MWh. Our arithmetic on a 100 MW hall at full load for a year: about 171,000 gallons a day at Microsoft’s fleet WUE, or about 1.33 million gallons a day at the Singapore median. Who pays is the water utility and the operator that holds the permit. The stuck step is wet evaporative cooling versus dry or closed-loop designs, plus local water stress — which is why Singapore is studying seawater cooling for Jurong Island. Why an investor cares: water is becoming a second scarce input beside power. A hall that cannot get a water budget or a discharge permit is as stuck as one waiting for a grid connection. This page builds on [what an AI data center is](https://www.sotkn.com/explained/what-is-an-ai-data-center/) and separates fleet averages from the local litre that actually blocks a build. ## How much water does an AI data center use? There is no single number. Cooling design, climate and how the operator counts water all move the answer. The industry yardstick is water usage effectiveness (WUE): litres of on-site water for cooling and humidification, divided by the kilowatt-hours used by the IT gear. One litre per kilowatt-hour equals one cubic metre per megawatt-hour. Microsoft’s fully owned and controlled fleet averaged 0.27 L/kWh in FY25, down from 0.30 in FY24, with Americas at 0.34, Asia Pacific at 0.25 and EMEA at 0.03\. Google reports a consumptive (ISO Category 2) fleet WUE of 1.15 L/kWh for both 2023 and 2024, and says it used about 8.1 billion gallons across its own data centers and offices in 2024, up 28% year on year. Equinix’s global portfolio averaged 0.95 in 2024, or 1.55 at sites that use evaporative cooling. Singapore’s PUB puts the median data-centre water-efficiency indicator at 2.1 m³/MWh. | Figure | Value | What it measures | | ---------------------------------------------- | ------------------------------------------------------- | ------------------------------------------------------- | | **Microsoft owned fleet WUE, FY25** | 0.27 L/kWh (FY24: 0.30) | Withdrawal for humidification and cooling ÷ IT kWh | | **Microsoft Americas / APAC / EMEA WUE, FY25** | 0.34 / 0.25 / 0.03 L/kWh | Same method, by region | | **Google consumptive WUE, 2023–24** | 1.15 L/kWh | ISO Category 2 (freshwater input minus returned) | | **Google total water, 2024** | \~8.1 billion gallons | Data centers (excl. third-party) plus offices; +28% YoY | | **Google replenished, 2024** | \~4.5 billion gallons (\~64% of freshwater consumption) | Stewardship projects, not on-site cooling | | **Equinix portfolio WUE, 2024** | 0.95 (evaporative sites 1.55) | All IBX sites vs water-cooled sites only | | **Equinix water withdrawal, 2024** | 5,440 megalitres (\~1.44 billion gallons) | Global data center portfolio | | **Digital Realty NA colo WUE, 2024** | 14% better than 2023; 42% non-potable | Efficiency and source mix, not a single fleet WUE | | **PUB Singapore DC median WEI** | 2.1 m³/MWh | Sector benchmark (Dec 2025) | | **100 MW full-load day (our arithmetic)** | \~171k gal/day at 0.27; \~1.33M gal/day at 2.1 | WUE × 100 MW × 24 h, US gallons | *From Microsoft, Google, Equinix, Digital Realty, PUB and IEA disclosures listed under References. Fleet WUE numbers use different numerators; do not treat them as identical.* ## What does that mean in gallons per day? Convert carefully, and label the maths. A 100 megawatt IT hall running flat out for a year draws 876 gigawatt-hours of IT energy. At Microsoft’s FY25 fleet WUE of 0.27 L/kWh that is about 236 million litres, or roughly 62.5 million US gallons a year — about 171,000 gallons a day (our arithmetic). At Equinix’s evaporative-site WUE of 1.55 it is about 983,000 gallons a day. At Singapore’s 2.1 median it is about 1.33 million gallons a day. Real halls rarely sit at 100% IT load all year; the table is a scale check, not a site forecast. | 100 MW IT, full year at 100% load | At 0.27 L/kWh | At 0.95 L/kWh | At 2.1 L/kWh | | --------------------------------- | ------------------------------ | ------------------ | ------------------- | | IT energy | 876 GWh (100,000 kW × 8,760 h) | | | | Water volume | 236.5 million litres | 832 million litres | 1.84 billion litres | | US gallons / year | \~62.5 million | \~220 million | \~486 million | | US gallons / day | \~171,000 | \~602,000 | \~1.33 million | *Second Order arithmetic. Real halls rarely run at 100% IT load all year; the table shows the scale of the WUE gap, not a forecast for one site.* ![AI data center water usage effectiveness: Microsoft 0.27, Equinix 0.95, Google 1.15, Equinix evaporative 1.55, Singapore PUB median 2.1 litres per kilowatt-hour](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/10/ai-dc-water-wue-chart-1.png) Same megawatt, different water. Published WUE or WEI; Second Order arithmetic for bar heights. ## Why do the published numbers disagree? Because the numerator is not standardised. The IEA notes that Google’s WUE uses freshwater consumption (input minus water returned), while Microsoft, Amazon and Meta publish withdrawal-based measures that count water taken in, including water later discharged. Amazon has illustrated that converting a modelled 0.56 L/kWh consumption figure to withdrawal can raise it to 0.84 L/kWh for the same system. Compare fleets only after you know which definition you are reading. Absolute totals also mix boundaries. Google’s 8.1 billion gallons covers its own data centers and offices, not third-party colo halls. Equinix withdrew 5,440 megalitres in 2024 across its global portfolio — about 1.44 billion gallons (our arithmetic) — and consumed 4,180 megalitres. Meta says data centers drive the bulk of its water use and reports restoration volumes rather than a single headline fleet WUE in its 2025 sustainability report. ## Who pays for the water? - **The water utility or PUB.** They meter supply and set discharge rules. In Singapore, sites using at least 60,000 cubic metres a year must submit water-efficiency management plans, meter key areas and appoint a water-efficiency manager. - **The hyperscaler, when it owns the hall.** It chooses wet, dry or closed-loop cooling and carries the bill and the permit risk. Microsoft says its newer AI-optimised design uses chip-level liquid cooling in a closed loop and consumes zero water for cooling, avoiding an estimated 125,000 cubic metres a year per facility. - **The landlord, in colocation.** Operators such as Equinix and Digital Realty design the plant, report portfolio WUE and often pass water costs through to tenants. Digital Realty says 42% of its 2024 water needs came from non-potable sources and that its North America colocation WUE improved 14% year on year. See [what Equinix does](https://www.sotkn.com/stack/ai-data-centers/what-does-equinix-do/) and [what Digital Realty does](https://www.sotkn.com/stack/ai-data-centers/what-does-digital-realty-do/). - **The watershed and the permit desk.** In stressed basins, the scarce item is the right to take more water, not the price per cubic metre. Google says it chose air cooling at high water-risk sites including Waltham Cross in the United Kingdom, Mesa in Arizona and Canelones in Uruguay. ![Who pays for AI data center water: utility or PUB, hyperscaler or landlord, stuck step wet versus dry cooling, watershed and community](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/10/ai-dc-water-who-pays-1.png) Stuck step: wet evaporative cooling versus dry or closed-loop. Microsoft, Google, Equinix, PUB, IEA. ## What is the stuck step? Wet evaporative cooling versus dry or closed-loop designs — and the permit that locks the choice. Evaporative towers use water to dump heat with less electricity, which can improve power usage effectiveness (PUE) while raising WUE. Dry coolers and air cooling flip that trade-off. Closed-loop liquid cooling at the chip can cut evaporative water sharply, which is the path Microsoft describes for new AI halls. The IEA puts the wider water story in three buckets: about two-thirds of data-centre-related water sits upstream in electricity generation, about a quarter in direct on-site cooling, and the rest in chip manufacturing. On-site cooling is the part a landlord or hyperscaler can redesign. Grid water and chip water sit outside the building fence. Build timing still matters: a hall that cannot cool cannot open, just as one waiting for power cannot, which is why [how long it takes to build an AI data center](https://www.sotkn.com/explained/how-long-does-it-take-to-build-an-ai-data-center/) and water permits now sit on the same critical path. Our view: watch the local litre and the permit language, not only the fleet WUE press line. A 0.27 fleet average does not unlock a new hall in a basin that has already said no. ## How does Asia constrain the litre? - **Singapore: benchmarks and paperwork.** PUB’s median data-centre WEI is 2.1 m³/MWh. Large users file annual plans. Freshwater is scarce by design, so EDB and JTC have flagged seawater-based cooling research with NUS for the Jurong Island low-carbon data centre park. Landlords such as Keppel DC REIT and the operators named in recent power allocations still have to clear water and discharge rules beside power. See [what Keppel DC REIT does](https://www.sotkn.com/stack/ai-data-centers/what-does-keppel-dc-reit-do/) and our [Singapore Pulse](https://www.sotkn.com/markets/singapore/singapore-data-centre-licence-october-2026/). - **Malaysia and Johor: growth next door.** Cross-border AI and cloud builds pull on Johor’s power and water systems even when the customer sits in Singapore. Our [Malaysia Pulse](https://www.sotkn.com/markets/malaysia/malaysia-data-centre-power-use-october-2026/) tracks connected power; water permits are the quieter twin constraint. - **Australia: landlords under two books.** NEXTDC and peers still design for local climate and utility rules as they book hyperscale halls. See [what NEXTDC does](https://www.sotkn.com/stack/ai-data-centers/what-does-nextdc-do/). ## What would prove this wrong? - **Fleet WUE stops falling while absolute water keeps rising fast.** Then efficiency claims are not offsetting AI load growth — Google’s +28% absolute water in 2024 is the pattern to watch. - **New AI halls keep choosing evaporative cooling in high-stress basins.** That would mean power scarcity is still beating water scarcity in siting decisions. - **Singapore’s median WEI falls toward 1.0 without seawater projects.** That would show freshwater efficiency alone can close the gap. ## Practical takeaways - **Ask for the numerator.** Withdrawal and consumption WUE are not the same number. - **Scale the hall, not the press release.** Convert WUE × IT megawatts × hours into gallons per day for the site you care about. - **Read the permit with the PUE.** A low-PUE wet design can still be blocked where water is scarce. - **Separate restoration from reduction.** Google’s 4.5 billion gallons replenished and Meta’s 1.59 billion gallons restored are stewardship projects, not proof the hall used less water. - **In Asia, check PUB and discharge rules early.** Power allocations do not replace water-efficiency plans. ## Common misconceptions - **People assume every AI data center uses millions of gallons a day.** Actually a closed-loop or dry-cooled hall can use far less on site; Microsoft’s new design targets zero cooling water. - **People assume a lower WUE always means a greener hall.** Actually dry cooling can raise electricity use; IEA treats energy and water as a joint trade-off. - **People assume replenishment cancels the local withdrawal.** Actually restoration projects often sit in other watersheds or later years; the local metre still runs. ## FAQ **How much water does an AI data center use?** It depends on the cooling design and the climate. Operators report water usage effectiveness (WUE) in litres per kilowatt-hour of IT energy. Microsoft’s owned fleet averaged 0.27 L/kWh in FY25\. Google’s consumptive fleet WUE was 1.15 L/kWh in 2023 and 2024\. Singapore’s PUB median for data centres is 2.1 m³/MWh. A 100 MW hall at full load for a year would use about 171,000 gallons a day at Microsoft’s fleet WUE, or about 1.33 million gallons a day at the Singapore median (our arithmetic). **What is water usage effectiveness (WUE)?** WUE is annual water used for cooling and humidification, divided by the kilowatt-hours used by IT equipment, usually in litres per kilowatt-hour. One L/kWh equals one m³/MWh. Operators disagree on whether the numerator is withdrawal (water taken in) or consumption (water not returned, mostly evaporated), so two published WUE numbers are not always comparable. **Why do AI data centers use water?** Servers make heat. Evaporative cooling and cooling towers use water because evaporating a litre of water removes a lot of heat with less electricity than dry chillers in many climates. Closed-loop liquid cooling at the chip can cut or eliminate that evaporative water, at the cost of different capital and power. **Who pays for the water?** The site pays the water utility or, in Singapore, PUB, for metered supply and for trade-effluent discharge rules. The hyperscaler pays if it owns the hall; the landlord pays or passes the cost through if it is colocation. In water-stressed places, the community and the permit authority set how much new load can connect. **Is dry cooling always better?** Not on every score. Dry or air cooling uses little or no water on site, but often uses more electricity for the same cooling load, which can raise PUE. IEA notes the trade-off explicitly. The stuck step is local: where water is scarce, dry or closed-loop wins; where water is abundant and power is tight, wet cooling can still be chosen. **How does Singapore constrain data centre water?** PUB’s median water efficiency for data centres is 2.1 m³/MWh. Sites using at least 60,000 m³ a year must submit water-efficiency plans, meter key areas and appoint a water-efficiency manager. EDB and JTC have also pointed to seawater-based cooling research for Jurong Island because of climate and freshwater constraints. **Is this a list of stocks to buy or sell?** No. Second Order Explained pages explain how the money and the constraints move. They do not give buy or sell calls. Related: [what is an AI data center](https://www.sotkn.com/explained/what-is-an-ai-data-center/); [what is a hyperscaler](https://www.sotkn.com/explained/hyperscaler-explained/); [what Digital Core REIT does](https://www.sotkn.com/stack/ai-data-centers/what-does-digital-core-reit-do/); [what Applied Digital does](https://www.sotkn.com/stack/ai-data-centers/what-does-applied-digital-do/). ## Stack doors Layer: [AI Data Centers](https://www.sotkn.com/stack/ai-data-centers/). Also: [Hyperscalers](https://www.sotkn.com/stack/hyperscalers/), [Neoclouds](https://www.sotkn.com/stack/neoclouds/), [Semiconductors](https://www.sotkn.com/stack/semiconductors/). Markets: [US](https://www.sotkn.com/markets/us/), [Singapore](https://www.sotkn.com/markets/singapore/), [Malaysia](https://www.sotkn.com/markets/malaysia/), [Australia](https://www.sotkn.com/markets/australia/). [Get the Wire](https://www.sotkn.com/#/portal/signup) — short daily notes when data center power, water and build dates move. **How we check this** WUE, absolute water volumes, replenishment and Singapore benchmarks come from Microsoft’s datacenter efficiency page and 2025 sustainability report, Google’s 2025 Environmental Report and its AI-impact paper, Meta’s 2025 Sustainability Report, Equinix’s 2024 Sustainability Data Summary, Digital Realty’s 2024 Impact Report, PUB’s sectoral benchmark, the IEA’s Powering Artificial Intelligence paper, and Singapore EDB’s Jurong Island note — all listed under References and re-checked on 16 Oct 2026\. Where we convert units or scale a 100 MW example, it is labelled “our arithmetic”. WUE numerators differ (withdrawal versus consumption); we state which each operator uses and do not average them. This page explains how water and money move. It does not give buy or sell calls. **Last reviewed:** 16 Oct 2026 · **Next review:** after the next Google, Microsoft and Equinix water or sustainability disclosures Spot an error? [Tell us](mailto:sean@thekopinotes.com) and we will correct it and note the change here. [Our editorial standards →](https://www.sotkn.com/editorial-standards/) ## References 1. [Microsoft Datacenters — Measuring energy and water efficiency (WUE/PUE tables, FY24–FY25)](https://datacenters.microsoft.com/sustainability/efficiency/?ref=sotkn.com) — global and regional WUE; definition of WUE as litres for humidification and cooling per IT kWh 2. [Microsoft — 2025 Environmental Sustainability Report](https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/msc/documents/presentations/CSR/2025-Microsoft-Environmental-Sustainability-Report.pdf?ref=sotkn.com) — zero-water-for-cooling AI datacenter design; \~125,000 m³ avoided per facility per year; closed-loop chip-level liquid cooling; water-positive 2030 3. [Google — 2025 Environmental Report](https://www.gstatic.com/gumdrop/sustainability/google-2025-environmental-report.pdf?ref=sotkn.com) — 8.1 billion gallons total water use in 2024 across data centers and offices; +28% YoY; 4.5 billion gallons replenished; 72% freshwater withdrawals from low-risk sources; air cooling at high water-risk sites 4. [Google — Measuring the environmental impact of delivering AI at Google Scale (2025)](https://arxiv.org/html/2508.15734?ref=sotkn.com) — fleet WUE Category 2 (consumptive) of 1.15 L/kWh for 2023 and 2024 5. [Meta — 2025 Sustainability Report](https://sustainability.atmeta.com/wp-content/uploads/2025/08/Meta%5F2025-Sustainability-Report.pdf?ref=sotkn.com) — data centers drive bulk of water use; water-positive 2030; 1.59 billion gallons restored in high/medium stress regions in 2024 6. [Equinix — 2024 Sustainability Data Summary (April 2025)](https://www.equinix.com/content/dam/eqxcorp/en%5Fus/documents/resources/data-sheets/ds%5Fsustainability%5Fdata%5Fen.pdf?ref=sotkn.com) — withdrawal 5,440 ML; consumption 4,180 ML; portfolio WUE 0.95; evaporative-site WUE 1.55; 37% non-potable 7. [Digital Realty — 2024 Impact Report](https://investor.digitalrealty.com/static-files/2fb13aa7-f49d-4528-8599-9a52810911d1?ref=sotkn.com) — 14% WUE improvement in North America colocation; 42% non-potable water; dry cooling designs; Ecolab conservation pilot 8. [PUB Singapore — Sectoral Water Efficiency Benchmark (Dec 2025)](https://www.pub.gov.sg/Public/WaterLoop/Water-Conservation/Resources-on-Water-Efficiency-Measures/Sectoral-Water-Efficiency-Benchmark?ref=sotkn.com) — data centres median WEI 2.1 m³/MWh 9. [IEA — Powering Artificial Intelligence](https://iea.blob.core.windows.net/assets/cd360a30-aaef-43cd-84e6-a125c3a0bc67/PoweringArtificialIntelligence.pdf?ref=sotkn.com) — water/energy cooling trade-off; \~two-thirds of DC-related water upstream in electricity generation; WUE numerator differences across hyperscalers 10. [EDB Singapore — Jurong Island low-carbon data centre park (green energy testbed)](https://www.edb.gov.sg/en/news-and-insights/data-centre-testbed-powered-by-green-energy-among-new-projects-to-fuel-jurong-islands-green-push?ref=sotkn.com) — JTC and NUS study on seawater-based cooling because of climate and water constraints S **[Sean](https://www.sotkn.com/author/sean/)** Writes Second Order by TKN — a plain-English map of AI infrastructure and semiconductors for non-specialist investors. Focus: who gets paid, where money sticks, and which step is stuck. Covers AI infrastructure, semiconductors, and second-order stack reads. [Author page →](https://www.sotkn.com/author/sean/) · [Editorial standards →](https://www.sotkn.com/editorial-standards/)