AI water and energy use: the disclosed numbers
How much water and electricity does AI use? Every figure here comes from a company report, a research paper or an IEA or Berkeley Lab study, with a link to the page it is printed on. Withdrawal, consumption, per prompt and per company are kept in separate columns.
0.26 mL
Water per median Gemini Apps text prompt
Google technical paper, August 2025, using May 2025 data. On-site consumption, inference only; 0.24 Wh of energy.
Link0.34 Wh
Energy per average ChatGPT query, as stated by OpenAI's CEO
Sam Altman blog post, June 2025. Water stated as 0.000085 gallons (about 0.32 mL). No methodology published.
Link41.1M m³
Google water consumption in 2025 (10,869 million gallons)
41.1 million m³. Google 2026 Environmental Report. Data centers and offices; up 34% on 2024.
Link13.3M m³
Microsoft water withdrawal in FY2025
13.3 million m³. Microsoft 2026 Environmental Data Fact Sheet. Consumption was 8.2 million m³.
Link415 TWh
Global data centre electricity use in 2024, about 1.5% of the world total
IEA, Energy and AI (April 2025). All data centres, not only AI.
Link945 TWh
IEA Base Case projection for global data centre electricity in 2030
Just under 3% of global electricity consumption in 2030.
Link192 TWh
US data center electricity use in 2024, 4.7% of US consumption
Lawrence Berkeley National Laboratory, 2025 Update (June 2026). Reference Case: 649 TWh by 2030.
Link560B L
Global data centre water use per year, incl. power generation and chips
560 billion litres. IEA, Energy and AI. About one quarter is direct on-site cooling.
Link
Arguments about how much water AI uses rarely line up because the numbers measure different things. A company can report withdrawal (all water taken in) or consumption (water that evaporates or is otherwise not returned). It can count only on-site cooling water, or add the far larger volume used to generate the electricity its data centers draw. A per-prompt figure can cover only the chips doing the work, or the whole facility, or a full life cycle that includes manufacturing the servers. And a company total usually covers every data center and office, not just AI.
This dataset keeps those measures apart. The companies table lists what each company prints in its own sustainability report or data sheet, with units as stated, the scope of the figure, and a normalized value in million cubic meters or gigawatt-hours only where the conversion is unambiguous. The per prompt table lists every per-query disclosure and research estimate that could be checked against the original document, with a column saying exactly what each one covers. The grid table holds the macro estimates from the International Energy Agency and Lawrence Berkeley National Laboratory.
Every number links to the page or passage where it is printed. Where a company publishes nothing, as with OpenAI, Anthropic and xAI at the company level, the table says so rather than filling the gap with an outside estimate. Compare rows only after checking the scope and covers columns: a 45 mL life cycle figure and a 0.26 mL on-site figure are not measuring the same thing.
Company water and energy disclosures
Figures as printed in each company's own report, with units as stated, plus a normalized value where conversion is unambiguous. Scope differs by company: check the scope column before comparing rows. Empty cells mean the company does not state the figure.
| Company | Year | Scope | Water withdrawal (as stated) | Withdrawal (million m3) | Water consumption (as stated) | Consumption (million m3) | Electricity (as stated) | Electricity (GWh) | PUE | WUE (L/kWh) | Report | Notes |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Alibaba | FY2026 (April 2025 to March 2026) | Water and electricity: group. PUE and WUE: Alibaba Cloud self-built data centers | Not stated | Not stated | 17,464,161 m3 (labelled total water consumption) | 17.46 | 9,310,297 MWh purchased electricity | 9310.3 | 1.187 | 1.198 | Alibaba Group Environmental, Social and Governance Report 2026 (SEC Form 6-K exhibit 99.2) | Energy and PUE table on page 125, water on page 126, WUE history on page 49. WUE was 0.205 at mainly air-cooled sites and 1.799 at mainly evaporative sites. Leased data centers in China had PUE 1.224. The PDF text layer is font-encoded; values were decoded and checked against the surrounding tables. |
| Alibaba | FY2025 | Same as above | Not stated | Not stated | 13,479,552 m3 (labelled total water consumption) | 13.48 | 6,764,967 MWh purchased electricity | 6765 | 1.19 | 1.144 | Alibaba ESG Report 2026 (prior year column) | Not stated |
| Amazon | 2025 | Data centers only (leased, owned and shared) | 9.4 billion liters | 9.4 | Not stated | Not stated | Not stated | Not stated | 1.14 | 0.12 | 2025 Amazon Sustainability Report | Amazon reports withdrawal, not consumption, and does not publish data center electricity. WUE is liters withdrawn per kWh of IT load (WUE table on page 15; PUE on page 10). A company article puts 2025 withdrawal at about 2.5 billion gallons and says withdrawal at owned and operated sites fell 2% from 2024. |
| Amazon | 2024 | Data centers only | Not stated | Not stated | Not stated | Not stated | Not stated | Not stated | 1.15 | 0.15 | 2024 Amazon Sustainability Report | WUE on page 24, PUE on page 10. No withdrawal volume is given for 2024. |
| Anthropic | No public report | No sustainability report | No public report | Not stated | No public report | Not stated | No public report | Not stated | Not stated | Not stated | Form S-1 registration statement (October 26, 2026) | The S-1 discusses power and water for cooling as risk factors but discloses no water or electricity volumes. |
| Apple | FY2025 | Whole company corporate facilities (offices, retail, data centers, R&D) | 1,796 million gallons (total water use) | Not stated | Not stated | Not stated | 3,737,000 MWh | 3737 | Not stated | Not stated | Apple Environmental Progress Report 2026 | Apple reports total water use, which it does not split into withdrawal and consumption, so no normalized value is given (about 6.8 million m3 if converted). It includes 1,551 million gallons of freshwater plus recycled and other sources. Data centers and colocation used 2,585 million kWh of electricity in FY2025 (page 94). Corporate electricity is reported as 100% renewable electricity used (page 80). No PUE or WUE published. |
| Apple | FY2024 | Whole company corporate facilities | 1,756 million gallons (total water use) | Not stated | Not stated | Not stated | 3,777,000 MWh | 3777 | Not stated | Not stated | Apple Environmental Progress Report 2026 (prior year column) | Not stated |
| Baidu | 2025 | Baidu Core (water and electricity). PUE and WUE: self-built data centers | 2,183,887.2 tons | 2.18 | 1,547,476.4 tons | 1.55 | 1,225,921.5 MWh purchased electricity | 1225.9 | 1.2 | 1.65 | Baidu 2025 Environmental, Social and Governance Report | Energy, PUE and WUE on page 64, water on page 66. 1 ton of water is treated as 1 m3. |
| Baidu | 2024 | Baidu Core | 2,110,550.2 tons | 2.11 | 1,465,128.9 tons | 1.47 | 1,145,027.5 MWh purchased electricity | 1145 | Not stated | Not stated | Baidu 2025 ESG Report (prior year column) | Not stated |
| Digital Realty | 2025 | Global data center portfolio | Not stated | Not stated | Not stated | Not stated | Not stated | Not stated | 1.38 | 0.59 | Digital Realty press release on its 2025 Impact Report (May 27, 2026) | Taken from the company press release; the full Impact Report PDF on the investor site blocked automated access, so absolute water and electricity volumes were not verified. The release says water usage rose 3% from 2023 to 2025 while the portfolio grew 34%, and WUE improved 15.7% on the prior year. EMEA PUE was 1.31. |
| Equinix | 2025 | Whole data center portfolio | 6,040 ML | 6.04 | 4,600 ML | 4.6 | 8,300 GWh electricity | 8300 | 1.37 | 0.91 | Equinix 2025 Sustainability Report | PUE page 16, electricity page 18, water page 25, WUE page 26. WUE at sites using evaporative cooling was 1.41. Equinix says consumption rose 10% on 2024. Equinix updated its GHG boundary in 2025, so the electricity figure is not directly comparable with the 2024 energy total. |
| Equinix | 2024 | Whole data center portfolio | 5,440 ML | 5.44 | 4,180 ML | 4.18 | 8,870 GWh total energy (Scope 1 and 2 sources) | Not stated | 1.39 | 0.95 | Equinix 2024 Sustainability Data Summary | Energy and PUE on page 9. WUE at evaporative-cooled sites was 1.55. |
| 2025 | Whole company (data centers and offices) | 14,689 million gallons | 55.6 | 10,869 million gallons | 41.14 | 43,586,600 MWh | 43586.6 | 1.09 | Not stated | Google 2026 Environmental Report | Data centers alone in 2025: withdrawal 13,562 million gallons, consumption 10,523 million gallons (39.83 million m3), electricity 42,415,800 MWh. Consumption rose 34% on 2024. PUE is the fleet-wide average for Google-owned data center campuses (page 95); electricity table on page 93. Google does not publish a fleet WUE in this report. | |
| 2024 | Whole company (data centers and offices) | 11,011 million gallons | 41.68 | 8,135 million gallons | 30.79 | 31,713,900 MWh | 31713.9 | 1.09 | 1.15 | Google 2026 Environmental Report (five-year table) | Data center electricity in 2024: 30,637,100 MWh. WUE 1.15 L/kWh is the consumptive (ISO Category 2) value for 2023 and 2024 stated in Google's Gemini paper (arXiv 2508.15734, page 6), not in the environmental report. | |
| IBM | 2025 | Data centers (PUE only) | Not stated | Not stated | Not stated | Not stated | Not stated | Not stated | 1.39 | Not stated | IBM energy and emissions page | IBM states the 2025 PUE is an estimated weighted average, a 28.0% cooling efficiency improvement on 2019, and that total energy fell 8.0% versus 2024. No water or electricity volumes on the page. |
| Meta | 2024 | Whole company (data centers and offices) | 5,637 ML | 5.637 | 3,123 ML | 3.123 | 18,423,634 MWh | 18423.6 | 1.08 | 0.19 | Meta 2025 Environmental Data Index | Data centers alone: withdrawal 4,145 ML, consumption 2,974 ML, electricity 18,061,781 MWh (page 6). WUE is withdrawal in liters divided by IT electricity in kWh (page 10). A further 1,019 ML withdrawn for data center construction is excluded. Latest Meta data published as of October 2026 covers 2024. |
| Meta | 2023 | Whole company (data centers and offices) | 5,274 ML | 5.274 | 3,078 ML | 3.078 | 15,325,314 MWh | 15325.3 | 1.08 | 0.18 | Meta 2025 Environmental Data Index | Data centers alone: withdrawal 3,881 ML, consumption 2,938 ML, electricity 14,975,435 MWh. |
| Microsoft | FY2025 (July 2024 to June 2025) | Whole company (offices, datacenters, other facilities) | 13,266 ML | 13.266 | 8,170 ML | 8.17 | 37,026,353 MWh | 37026.4 | 1.17 | 0.27 | 2026 Environmental Data Fact Sheet; 2026 Environmental Sustainability Report | Water table on fact sheet page 7, electricity on page 6. PUE 1.17 (report page 23) and WUE 0.27 L/kWh (report page 24) are global averages for owned datacenters; Microsoft defines WUE as datacenter water for cooling and humidification divided by energy use. Prior years were restated in this edition. Microsoft says it replenished more than 14.2 million m3 in FY25, exceeding withdrawals. |
| Microsoft | FY2024 (July 2023 to June 2024) | Whole company (offices, datacenters, other facilities) | 11,590 ML | 11.59 | 6,693 ML | 6.693 | 29,829,540 MWh | 29829.5 | Not stated | Not stated | 2026 Environmental Data Fact Sheet (restated prior year) | Restated values as printed in the 2026 fact sheet; they differ from figures first published for FY2024. |
| OpenAI | No public report | No sustainability report | No public report | Not stated | No public report | Not stated | No public report | Not stated | Not stated | Not stated | CEO blog post (per-query figure only) | OpenAI publishes no company water or energy totals. The only figure is a per-query estimate in a June 2025 blog post by its CEO; see the per-prompt table. |
| Oracle | FY2026 (June 2025 to May 2026) | Water: Oracle-owned real estate and owned data centers. Energy: Scopes 1 and 2 | 816,484,842 liters potable water use | Not stated | Not stated | Not stated | 8,789,499 MWh total energy use (not electricity only) | Not stated | Not stated | Not stated | Oracle Environmental and Social Impact FY26 Datasheet (October 2026) | Energy on page 3, water on page 4. Oracle reports potable water use (about 0.82 million m3), not withdrawal or consumption, and publishes no PUE or WUE. Total energy was 5,155,385 MWh in FY2025. The datasheet URL is updated in place with each edition. |
| Oracle | FY2025 | Same as above | 801,872,356 liters potable water use | Not stated | Not stated | Not stated | 5,155,385 MWh total energy use (not electricity only) | Not stated | Not stated | Not stated | Oracle Environmental and Social Impact FY26 Datasheet | Not stated |
| Tencent | 2025 | Water and electricity: group (headquarters and wholly owned subsidiaries). PUE and WUE: owned data centers | Not stated | Not stated | 12,151,404.1 tonnes (labelled water consumption) | 12.15 | 8,669,131.7 MWh | 8669.1 | 1.246 | 1.5 | Tencent Environmental, Social and Governance Report 2025 | Tencent labels this figure water consumption and says all water comes from municipal supply. 1 tonne of water is treated as 1 m3. PUE on page 4, WUE on page 47. |
| Tencent | 2024 | Same as above | Not stated | Not stated | 10,370,588.3 tonnes (labelled water consumption) | 10.37 | 6,429,609.9 MWh | 6429.6 | 1.257 | Not stated | Tencent ESG Report 2025 (prior year column) | Not stated |
| xAI (part of SpaceX since February 2026) | No public report | No sustainability report | No public report | Not stated | No public report | Not stated | No public report | Not stated | Not stated | Not stated | SpaceX Form S-1/A (June 3, 2026) | The filing discusses water scarcity as a risk to data center cooling but discloses no water or electricity volumes for xAI. |
Energy and water per prompt
Company disclosures and research estimates for a single AI request. They measure different things (median versus average, inference only versus life cycle, on-site water versus water embodied in hardware), so compare the covers column before comparing numbers.
| Organization | Model or service | Energy (Wh) | Water (mL) | Emissions (gCO2e) | What it covers | Source type | Date | Source | Notes |
|---|---|---|---|---|---|---|---|---|---|
| Gemini Apps, median text prompt (comprehensive method) | 0.24 | 0.26 | 0.03 | Inference only. Includes active accelerators, host CPU and RAM, idle machines and data center overhead. Water is on-site consumption, calculated as energy excluding data center overhead multiplied by the Google WUE of 1.15 L/kWh | Company disclosure (technical paper) | 2025-08-21 (data from May 2025) | arxiv.org | Also in the Google Cloud blog of August 22, 2025 and the Google 2026 Environmental Report. Excludes training, networking and user devices. Google says energy per median prompt fell 33x over the prior 12 months. | |
| Gemini Apps, median text prompt (narrow method, active accelerators only) | 0.1 | 0.12 | 0.02 | Inference only. Active TPU/GPU energy only, comparable to many earlier public estimates | Company disclosure (technical paper) | 2025-08-21 (data from May 2025) | arxiv.org | Table 2 of the paper and the Google Cloud blog give 0.10 Wh, 0.02 gCO2e and 0.12 mL. One sentence in the paper text instead gives 0.01 gCO2e and 0.15 mL for this method; this table uses the Table 2 values. | |
| Hugging Face (Delavande, Luccioni, Pierrard) | WAN2.1-T2V-14B, one short video clip | 109.47 | Not stated | Not stated | Inference only, measured on one H100 with CodeCarbon; GPU 93.83 + CPU 10.47 + RAM 5.17 Wh | Research blog (not peer-reviewed) | 2025-07-02 | huggingface.co | Clip of 81 frames at 480x832. Excludes data center overhead. |
| Hugging Face (Delavande, Luccioni, Pierrard) | AnimateDiff, one short video clip | 0.14 | Not stated | Not stated | Inference only, measured on one H100; GPU 0.11 + CPU 0.02 + RAM 0.01 Wh | Research blog (not peer-reviewed) | 2025-07-02 | huggingface.co | Clip of 16 frames at 512x512. Shows the roughly 800x spread between open video models. |
| Jegham et al. | DeepSeek-R1 on DeepSeek servers, long prompt (10k input, 1.5k output tokens) | 29.08 | over 200 | about 17 | Inference only, estimated. Water includes on-site and off-site (electricity generation) components | Preprint (arXiv v6) | 2025-11-24 | arxiv.org | The same model served on Azure was estimated at 7.41 Wh, about 34 mL and 2.5 gCO2e per long query. |
| Jegham et al. | GPT-5, long query with high reasoning effort | 33.8 | Not stated | Not stated | Inference only, estimated average | Preprint (arXiv v6) | 2025-11-24 | arxiv.org | A short query with minimal reasoning was estimated at 0.67 Wh. |
| Jegham et al. (University of Rhode Island and others) | GPT-4o, short prompt (100 input, 300 output tokens) | 0.42 | Not stated | Not stated | Inference only, estimated from public API throughput plus assumed hardware, PUE, WUE and grid factors | Preprint (arXiv v6) | 2025-11-24 | arxiv.org | Table 4 gives 0.423 plus or minus 0.085 Wh; text gives 0.42 Wh plus or minus 0.13 Wh, or 0.37 Wh without data center overhead. Title: How Hungry is AI? |
| Luccioni, Jernite and Strubell | Open text generation models, mean per query | 0.047 | Not stated | Not stated | Inference only, measured on A100 GPUs; reported as 0.047 kWh per 1,000 inferences | Peer-reviewed (ACM FAccT 2024) | 2024-06 | arxiv.org | Models tested are far smaller than current frontier chat models. |
| Luccioni, Jernite and Strubell (Hugging Face, Carnegie Mellon) | Open image generation models, mean per image | 2.907 | Not stated | Not stated | Inference only, measured on A100 GPUs; reported as 2.907 kWh per 1,000 inferences (standard deviation 3.31) | Peer-reviewed (ACM FAccT 2024) | 2024-06 | arxiv.org | The least efficient model tested used 11.49 kWh per 1,000 images (about 11.5 Wh each). Small open models, not production services. |
| Microsoft | Large production LLM, typical query (company blog on the Joule study) | 0.16 to 0.60 | 0.0 to 0.067 | Not stated | Inference only. On-site cooling water per query under conservative assumptions | Company disclosure (blog) | 2026-06-15 | microsoft.com | Microsoft describes the median water use as about one hundredth of a teaspoon. |
| Microsoft Research | Frontier-scale LLM, typical query under optimized production serving | 0.31 | Not stated | Not stated | Inference only. Bottom-up model of energy per query; interquartile range 0.16 to 0.60 Wh | Peer-reviewed (Joule) | 2026-04 | microsoft.com | Oviedo et al., Joule, doi 10.1016/j.joule.2026.102430. Summarized in a Nature Energy research highlight on May 27, 2026 (https://www.nature.com/articles/s41560-026-02084-9). Longer reasoning or agentic queries use more. |
| Mistral AI | Le Chat, one 400-token response (Mistral Large 2) | Not stated | 45 | 1.14 | Marginal inference impact from a life cycle assessment, including upstream server manufacturing, excluding user devices | Company disclosure (LCA with Carbone 4 and ADEME, peer-reviewed by Resilio and Hubblo) | 2025-07-22 | mistral.ai | No energy per prompt disclosed. The same study puts training plus 18 months of use of Mistral Large 2 at 20.4 ktCO2e and 281,000 m3 of water (as of January 2025). Water figure includes water embodied in hardware, so it is not comparable with on-site cooling figures. |
| OpenAI | ChatGPT, average query | 0.34 | 0.32 | Not stated | Not stated. CEO blog statement without published methodology | CEO blog statement (no methodology) | 2025-06 | blog.samaltman.com | Stated as about 0.000085 gallons of water, roughly one fifteenth of a teaspoon; 0.32 mL is our unit conversion. Model, scope (on-site or including power generation water) and date range are not specified. |
Data center electricity and water: global and US estimates
Macro estimates from the International Energy Agency and Lawrence Berkeley National Laboratory. These cover all data centers, not only AI.
| Source | Geography | Metric | Year | Value | Unit | Share of demand | Estimate or projection | Source link | Notes |
|---|---|---|---|---|---|---|---|---|---|
| IEA Energy and AI (April 2025) | World | Data centre electricity consumption | 2024 | 415 | TWh | about 1.5% of global electricity | Estimate | iea.blob.core.windows.net | United States 45%, China 25%, Europe 15% of the 2024 total. Growth of about 12% a year since 2017. |
| IEA Energy and AI (April 2025) | World | Data centre electricity consumption, Base Case | 2030 | 945 | TWh | just under 3% of global electricity | Projection | iea.blob.core.windows.net | About 15% a year growth from 2024 to 2030. Electricity use in accelerated (mainly AI) servers grows about 30% a year, conventional servers about 9%. |
| IEA Energy and AI (April 2025) | World | Data centre electricity consumption, Base Case | 2035 | 1200 | TWh | Not stated | Projection | iea.blob.core.windows.net | Stated as around 1,200 TWh. Uncertainty widens after 2030; the range across IEA cases for 2035 is 700 to 1,700 TWh. |
| IEA Energy and AI (April 2025) | World | Data centre electricity consumption, Lift-Off Case | 2035 | over 1,700 | TWh | 4.4% of global electricity | Projection | iea.blob.core.windows.net | About 45% higher than the Base Case. |
| IEA Energy and AI (April 2025) | World | Data centre water consumption (direct cooling plus electricity generation plus chip manufacturing) | 2023 | 560 | billion litres | Not stated | Estimate | iea.blob.core.windows.net | About two-thirds from energy supply and electricity generation, one quarter from direct cooling, the rest from semiconductor manufacturing. Box 5.4. |
| IEA Energy and AI (April 2025) | World | Data centre water consumption, Base Case | 2030 | 1200 | billion litres | Not stated | Projection | iea.blob.core.windows.net | Stated as around 1,200 billion litres per year. |
| IEA Energy and AI (April 2025) | United States | Water consumption of a typical 100 MW hyperscale data centre | Current (report year) | 2 | million litres per day | Not stated | Estimate | iea.blob.core.windows.net | Total including indirect use; over 60% is indirect (electricity and supply chain). IEA equates it to about 6,500 households. |
| LBNL 2024 United States Data Center Energy Usage Report | United States | Data center electricity use | 2023 | 176 | TWh | 4.4% of US electricity | Estimate | eta-publications.lbl.gov | Superseded for 2024 onward by the 2025 update. |
| LBNL 2024 United States Data Center Energy Usage Report | United States | Data center electricity use, scenario range | 2028 | 325 to 580 | TWh | 6.7% to 12.0% of US electricity | Projection | eta-publications.lbl.gov | Equivalent to 74 to 132 GW of demand at 50% average utilization. |
| LBNL 2024 United States Data Center Energy Usage Report | United States | Direct (on-site) data center water consumption | 2023 | 66 | billion liters | Not stated | Estimate | eta-publications.lbl.gov | Hyperscale and colocation were 84% of the total. 21.2 billion liters in 2014. |
| LBNL 2024 United States Data Center Energy Usage Report | United States | Indirect water consumption from electricity generation for data centers | 2023 | nearly 800 | billion liters | Not stated | Estimate | eta-publications.lbl.gov | Equivalent to 4.52 L/kWh of indirect water on average. Roughly 12 times the direct on-site figure. |
| LBNL 2024 United States Data Center Energy Usage Report | United States | Direct water consumption, hyperscale data centers | 2028 | 60 to 124 | billion liters | Not stated | Projection | eta-publications.lbl.gov | Not stated |
| LBNL United States Data Center Energy Usage Report: 2025 Update | United States | Data center electricity use | 2024 | 192 | TWh | 4.7% of US electricity | Estimate | osti.gov | Published June 18, 2026 (doi 10.71468/P1RP4F). Excludes cryptocurrency mining. |
| LBNL United States Data Center Energy Usage Report: 2025 Update | United States | Data center electricity use, Reference Case | 2028 | 464 | TWh | Not stated | Projection | osti.gov | Not stated |
| LBNL United States Data Center Energy Usage Report: 2025 Update | United States | Data center electricity use, Reference Case | 2030 | 649 | TWh | 11.8% of US electricity | Projection | osti.gov | Would make data centers 33% of US load growth from 2024 to 2030. |
| LBNL United States Data Center Energy Usage Report: 2025 Update | United States | Data center electricity use, compounded uncertainty range | 2030 | 521 to 843 | TWh | 9.5% to 15.3% of US electricity | Projection | osti.gov | Not stated |
Methodology
How the data was compiled. On 2026-10-29 every figure was checked against the primary document: the company's own sustainability report, data sheet, regulatory filing or blog, the research paper itself, or the IEA and Berkeley Lab reports. Each link points to the PDF page (#page=N) or to a text fragment that appears verbatim on the cited web page. Figures that could only be found in news coverage, or that did not match the primary document, were left out.
Exceptions. Digital Realty's PUE and WUE come from the company's own press release about its 2025 Impact Report, because the report PDF could not be retrieved; this is noted in the row. Alibaba's ESG report PDF has a font-encoded text layer, so its values were decoded and cross-checked against neighbouring years and totals.
Definitions.
- Water withdrawal: all water taken from any source (municipal supply, surface water, groundwater) in the period (GRI 303-3).
- Water consumption: withdrawal minus water discharged back to the same source, mostly water evaporated in cooling (GRI 303-5).
- On-site versus indirect water: on-site water is used at the data center itself. Indirect water is used to generate the electricity it consumes; the IEA and Berkeley Lab estimate it is larger than on-site use.
- PUE (power usage effectiveness): total facility energy divided by IT equipment energy. 1.0 would mean no overhead.
- WUE (water usage effectiveness): liters of water per kilowatt-hour, usually of IT energy. Companies differ on whether the numerator is withdrawal or consumption and what it includes, so WUE values are not strictly comparable across companies.
Conversions. 1 US gallon = 3.785411784 liters. 1 megaliter (ML) = 1,000 m³. One tonne of water is treated as 1 m³. MWh are divided by 1,000 for GWh. Normalized columns are left empty when a company reports total energy rather than electricity, or a figure that is neither withdrawal nor consumption.
Corrections. If a figure is wrong, out of date, or a newer report has been published, email editors@thefrontierwire.com with the document link and page number. The table and the date above are updated.
Questions readers ask
How much water does ChatGPT use per prompt?
OpenAI has not published a measured figure. Its CEO, Sam Altman, wrote in June 2025 that an average ChatGPT query uses about 0.000085 gallons of water (about 0.32 mL) and 0.34 Wh of electricity, without describing the method. For comparison, Google measured 0.26 mL for a median Gemini Apps text prompt, Microsoft estimates 0.0 to 0.067 mL for a typical query to a large production model, and Mistral's life cycle assessment gives 45 mL for a 400-token Le Chat response, a figure that includes water used to make the hardware.
How much water does AI use in total?
No company reports AI separately from its other data center and office use. Company-wide water consumption was 41.1 million m³ for Google in 2025, 8.2 million m³ for Microsoft in FY2025 and 3.1 million m³ for Meta in 2024. The IEA estimates all data centres worldwide consume around 560 billion litres a year once electricity generation and chipmaking are included, rising to around 1,200 billion litres by 2030. Berkeley Lab puts direct on-site consumption by US data centers at 66 billion liters in 2023. OpenAI, Anthropic and xAI publish no company totals.
How much electricity does AI use?
The IEA estimates data centres used about 415 TWh in 2024, around 1.5% of global electricity, and projects around 945 TWh by 2030 in its Base Case, with electricity use in accelerated (mainly AI) servers growing about 30% a year. Berkeley Lab estimates US data centers used 192 TWh in 2024 (4.7% of US electricity) and projects 649 TWh by 2030 in its Reference Case. Per query, Google measured 0.24 Wh for a median Gemini text prompt and a Microsoft Research study in Joule estimates a median of 0.31 Wh for frontier-scale models.
What is WUE?
Water usage effectiveness is liters of water per kilowatt-hour of energy, usually IT energy. Lower is better. Reported values range from 0.12 L/kWh (Amazon, 2025) and 0.19 (Meta, 2024) to 0.27 (Microsoft, FY2025), 0.91 (Equinix, 2025), 1.15 (Google, consumptive basis), 1.198 (Alibaba Cloud, FY2026), 1.50 (Tencent, 2025) and 1.65 (Baidu, 2025). Companies define the numerator differently, so treat cross-company comparisons with care.
What is the difference between water withdrawal and water consumption?
Withdrawal is all the water a company takes in; consumption is the part not returned, mostly evaporated in cooling. Google withdrew 14,689 million gallons in 2025 and consumed 10,869 million gallons. Microsoft withdrew 13,266 megaliters in FY2025 and consumed 8,170. Amazon and Meta's WUE figures are based on withdrawal; Google's 1.15 L/kWh is based on consumption.
Sources
- Google 2026 Environmental Report
- Measuring the environmental impact of delivering AI at Google Scale (arXiv 2508.15734)
- Google Cloud blog: How much energy does Google's AI use? We did the math
- Microsoft 2026 Environmental Sustainability Report
- Microsoft 2026 Environmental Data Fact Sheet
- 2025 Amazon Sustainability Report
- 2024 Amazon Sustainability Report
- Meta 2025 Environmental Data Index
- Apple Environmental Progress Report 2026
- Oracle Environmental and Social Impact FY26 Datasheet
- IBM: Energy and emissions
- Equinix 2025 Sustainability Report
- Equinix 2024 Sustainability Data Summary
- Digital Realty press release: 2025 Impact Report (May 27, 2026)
- Tencent Environmental, Social and Governance Report 2025
- Alibaba Group ESG Report 2026 (SEC Form 6-K exhibit 99.2)
- Baidu 2025 ESG Report
- Anthropic, PBC Form S-1 (October 26, 2026)
- Space Exploration Technologies Corp. Form S-1/A (June 3, 2026)
- Sam Altman: The Gentle Singularity
- Mistral AI: Our contribution to a global environmental standard for AI
- Oviedo et al., Energy use of AI inference, efficiency pathways, and test-time scaling (Joule, 2026)
- Microsoft Cloud blog: Scaling AI with 8 to 20x energy efficiency
- Nature Energy research highlight: Energy use in AI responses
- Jegham et al., How Hungry is AI? (arXiv 2505.09598v6)
- Luccioni, Jernite and Strubell, Power Hungry Processing (ACM FAccT 2024)
- Hugging Face: How Much Power does a SOTA Open Video Model Use?
- IEA, Energy and AI (April 2025)
- LBNL, 2024 United States Data Center Energy Usage Report
- LBNL, United States Data Center Energy Usage Report: 2025 Update
Cite this dataset
The Frontier Wire. (2026). AI water and energy use: the disclosed numbers [Data set]. https://thefrontierwire.com/data/ai-water-and-energy-use/. Updated October 29, 2026.
Link to it
<a href="https://thefrontierwire.com/data/ai-water-and-energy-use/">AI water and energy use: the disclosed numbers</a> (The Frontier Wire, updated October 29, 2026)