Open evidence atlas · v0.2
See what AI water headlines leave out.
Computers produce heat. Cooling and electricity production can consume water in different places. Manufacturing is another boundary. There is no universal water-per-prompt figure and no verified global AI water total in this release.
Explore company reporting, regional context and conditional choices. Company figures are self-reported; selected metrics have limited assurance. ThirstyAI has not independently remeasured them.
Spot the missing boundary
Compare the period, fleet and withdrawal/consumption definition in the company disclosures below. JavaScript enables the interactive claim challenge.
Company water disclosures
Different periods, units and fleets: compare reporting, not company performance. Null is not zero.
| Company / metric | Value / unit | Period / geography / scope | Boundary / confidence | Source / locator / caveat |
|---|---|---|---|---|
| Google / withdrawal | 14689.0 million gallons (as reported) | 2025-01-01/2025-12-31 / Global / Global operations | withdrawal / not_applicable / extraction High | Google / GOOGLE26 PDF p.96 Water use by data centers and offices, 2025; water methodology pp.87–88 / Not AI-only. Reported water excludes seawater and brackish water. Consumption = withdrawal minus discharge; some values may be estimated. Years and fleets differ. |
| Google / consumption | 10869.0 million gallons (as reported) | 2025-01-01/2025-12-31 / Global / Global operations | consumption / not_applicable / extraction High | Google / GOOGLE26 PDF p.96 Water use by data centers and offices, 2025; water methodology pp.87–88 / Not AI-only. Reported water excludes seawater and brackish water. Consumption = withdrawal minus discharge; some values may be estimated. Years and fleets differ. |
| Google / withdrawal | 13562.0 million gallons (as reported) | 2025-01-01/2025-12-31 / Global / Operational data centers: owned and fully leased | withdrawal / not_applicable / extraction High | Google / GOOGLE26 PDF p.96 Water use by data centers and offices, 2025; water methodology pp.87–88 / Not AI-only. Reported water excludes seawater and brackish water. Consumption = withdrawal minus discharge; some values may be estimated. Years and fleets differ. |
| Google / consumption | 10523.0 million gallons (as reported) | 2025-01-01/2025-12-31 / Global / Operational data centers: owned and fully leased | consumption / not_applicable / extraction High | Google / GOOGLE26 PDF p.96 Water use by data centers and offices, 2025; water methodology pp.87–88 / Not AI-only. Reported water excludes seawater and brackish water. Consumption = withdrawal minus discharge; some values may be estimated. Years and fleets differ. |
| Google / PUE | 1.09 facility kWh / IT kWh | 2025-01-01/2025-12-31 / Global / Data-center fleet | not_applicable / not_applicable / extraction High | Google / GOOGLE26 PDF p.15, fleet-wide average PUE / Not AI-only; reporting periods, fleets and water definitions differ. |
| Google / WUE | Not extracted not extracted | 2025-01-01/2025-12-31 / Global / Latest annual report | not_applicable / not_applicable / extraction High | Google / GOOGLE26 Report searched for WUE; no compatible fleet value located / No numeric value extracted from this annual report. Do not substitute the older Gemini serving-study WUE. |
| Microsoft / withdrawal | 13266.0 ML | FY2025 ended 2025-06-30 / Global / Enterprise offices, owned/leased datacenters and labs | withdrawal / not_applicable / extraction High | Microsoft / MS26DATA PDF p.7, Table 8, FY25 / Not AI-only; reporting periods, fleets and water definitions differ. |
| Microsoft / consumption | 8170.0 ML | FY2025 ended 2025-06-30 / Global / Enterprise offices, owned/leased datacenters and labs | consumption / not_applicable / extraction High | Microsoft / MS26DATA PDF p.7, Table 8, FY25 / Not AI-only; reporting periods, fleets and water definitions differ. |
| Microsoft / PUE | 1.17 facility kWh / IT kWh | FY2025 ended 2025-06-30 / Global / Data-center global average | not_applicable / not_applicable / extraction High | Microsoft / MS26REPORT PDF p.23, Datacenters energy / Not AI-only; reporting periods, fleets and water definitions differ. |
| Microsoft / WUE | 0.27 L/kWh (publisher definition) | FY2025 ended 2025-06-30 / Global / Owned datacenters | water_use_unspecified / energy_use_unspecified / extraction High | Microsoft / MS26REPORT PDF p.24; definition endnote 12 p.64 / Inspected 2026 definition says cooling/humidification water use divided by energy, without explicitly resolving withdrawal vs consumption and IT denominator. Not accepted as consumptive WUE by the calculator. |
| Meta / withdrawal | 5637.0 ML | 2024-01-01/2024-12-31 / Global / Operations excluding separately disclosed construction | withdrawal / not_applicable / extraction High | Meta / META25 PDF p.9, section 3.1, 2024 column / Not AI-only; reporting periods, fleets and water definitions differ. |
| Meta / consumption | 3123.0 ML | 2024-01-01/2024-12-31 / Global / Operations | consumption / not_applicable / extraction High | Meta / META25 PDF p.11, section 3.2, 2024 column / Not AI-only; reporting periods, fleets and water definitions differ. |
| Meta / withdrawal | 4145.0 ML | 2024-01-01/2024-12-31 / Global / Data centers | withdrawal / not_applicable / extraction High | Meta / META25 PDF p.9, section 3.1, 2024 column / Not AI-only; reporting periods, fleets and water definitions differ. |
| Meta / consumption | 2974.0 ML | 2024-01-01/2024-12-31 / Global / Data centers | consumption / not_applicable / extraction High | Meta / META25 PDF p.11, section 3.2, 2024 column / Not AI-only; reporting periods, fleets and water definitions differ. |
| Meta / PUE | 1.08 facility kWh / IT kWh | 2024-01-01/2024-12-31 / Global / Data centers | not_applicable / not_applicable / extraction High | Meta / META25 PDF p.8, section 2.4, 2024 column / Not AI-only; reporting periods, fleets and water definitions differ. |
| Meta / WUE | Not extracted L / IT kWh | 2024-01-01/2024-12-31 / Global / Data centers | withdrawal / IT_kWh / extraction High | Meta / META25 PDF p.18, methodology PUE/WUE / Definition is withdrawal-based. No fleet numeric WUE located in inspected current report/index; preserve missing value. |
Location context, not a hotspot ranking
7 curated published location records; 5 US state-context matches; 0 verified exact facility/basin stress joins. No coordinates or water-supply basins are guessed. WRI administrative scores use a 1979–2019 modeled baseline. Unknown is not low risk.
| Published location / water | Regional context | Source trail |
|---|---|---|
| Google Henderson, Nevada / 239.4 million gallons (as reported) / consumption / 2025 | Nevada / High (40-80%) / score 3.569639288 dimensionless / 1979-2019 baseline; Aqueduct 4.0 release 2023 / geographic confidence Medium | Google / GOOGLE26 PDF p.97 Water use by data center location, Henderson NV World Resources Institute / WRIADMIN province_baseline row 6456; name_1=Nevada, indicator_name=bws, weight=Tot |
| Google Mesa, Arizona / 9.7 million gallons (as reported) / consumption / 2025 | Arizona / Extremely High (>80%) / score 4.392122413 dimensionless / 1979-2019 baseline; Aqueduct 4.0 release 2023 / geographic confidence Medium | Google / GOOGLE26 PDF p.97 Water use by data center location, Mesa AZ World Resources Institute / WRIADMIN province_baseline row 2284; name_1=Arizona, indicator_name=bws, weight=Tot |
| Microsoft Phoenix, Arizona / 981 ML / withdrawal / FY2025 ended 2025-06-30 | Arizona / Extremely High (>80%) / score 4.392122413 dimensionless / 1979-2019 baseline; Aqueduct 4.0 release 2023 / geographic confidence Medium | Microsoft / MS26DATA PDF p.25 Table 15 Phoenix AZ; exclusions p.26 World Resources Institute / WRIADMIN province_baseline row 2284; name_1=Arizona, indicator_name=bws, weight=Tot |
| Meta Los Lunas, New Mexico / 252 ML / withdrawal / 2024 | New Mexico / Extremely High (>80%) / score 4.324930145 dimensionless / 1979-2019 baseline; Aqueduct 4.0 release 2023 / geographic confidence Medium | Meta / META25 PDF p.9 Water withdrawal by facility, Los Lunas NM World Resources Institute / WRIADMIN province_baseline row 2437; name_1=New Mexico, indicator_name=bws, weight=Tot |
| Meta Mesa, Arizona / 57 ML / withdrawal / 2024 | Arizona / Extremely High (>80%) / score 4.392122413 dimensionless / 1979-2019 baseline; Aqueduct 4.0 release 2023 / geographic confidence Medium | Meta / META25 PDF p.9 Water withdrawal by facility, Mesa AZ World Resources Institute / WRIADMIN province_baseline row 2284; name_1=Arizona, indicator_name=bws, weight=Tot |
WRI Aqueduct extract, CC BY 4.0. State proxy is not current facility stress or evidence of causal harm.
India groundwater: all users, not AI extraction
36 state/UT observations from the 2025 CGWB assessment. Stage of extraction is annual extraction / annual extractable groundwater, not stored groundwater depleted. Ministry of Jal Shakti / PIB / CGWB25. WRI India / WRIINDIA26 separately studies data-centre siting and stress; we do not recreate its commercial location inventory.
Download the attributed state tableDecision lab
ILLUSTRATIVE INPUTS - NOT MEASUREMENTSHold useful IT work constant and compare direct consumptive WUE, PUE and electricity-related consumption. Manufacturing and local scarcity impacts are excluded. JavaScript enables the controls; these are hypothetical inputs, not measured company defaults.
Conditional futures to 2035
ILLUSTRATIVE INPUTS - NOT MEASUREMENTSDemand growth and efficiency rates are assumptions. This is not a forecast of global AI water, drought or achieved savings.
Numerical claim register
| Claim / value | Period / geography | Boundary / type / confidence | Source / locator / caveat |
|---|---|---|---|
| Data-centre electricity demand, 2025 / 485 TWh/year | 2025 / Global | All data centres, not AI alone / not_applicable / IEA estimate / High | International Energy Agency / IEA26 Executive summary, updated demand outlook paragraph / Electricity, not water; report vintage 2026. |
| Data-centre electricity demand, 2030 / 950 TWh/year | 2030 / Global | All data centres, not AI alone / not_applicable / IEA projection / Medium | International Energy Agency / IEA26 Executive summary, updated demand outlook paragraph / Conditional central projection; not a measured future outcome. |
| India groundwater stage of extraction / 60.63 % | 2025 / India | Annual extraction divided by annual extractable groundwater; all uses / not_applicable / Government reported / High | Ministry of Jal Shakti / PIB / CGWB25 Main text and Annexure I, Grand Total / Not the percentage of stored groundwater depleted. Does not attribute extraction to AI. |
| Median Gemini Apps text-prompt serving energy / 0.24 Wh/prompt | May 2025 / Google serving fleet | Company-authored serving methodology, including facility overhead / not_applicable / Production study / High | Elsworth et al. / Google / GOOGLE25 Table 2, comprehensive approach / Not every prompt or model; do not apply PUE again to facility-energy metric. |
| Median Gemini Apps text-prompt cooling water / 0.26 mL/prompt | May 2025 / Google serving fleet | Allocated direct consumptive cooling water; previous-year WUE / consumption / Company-modeled allocation / Medium | Elsworth et al. / Google / GOOGLE25 Section 3.3 and Table 2 / Not electricity-related or full lifecycle water. Cannot be generalized to all AI tasks. |
| US data-centre electricity reference case / 649 TWh/year | 2030 / United States | All US data centres; 2025 Update published June 2026 / not_applicable / LBNL projection / Medium | Lawrence Berkeley National Laboratory / LBNL26 Official publication abstract, Reference Case / Different model from IEA; not a water total. Do not combine projections as observations. |
| Google withdrawal / 14689 million gallons (as reported) | 2025-01-01/2025-12-31 / Global | Global operations / withdrawal / company_reported / High | Google / GOOGLE26 PDF p.96 Water use by data centers and offices, 2025; water methodology pp.87–88 / Not AI-only. Reported water excludes seawater and brackish water. Consumption = withdrawal minus discharge; some values may be estimated. Years and fleets differ. |
| Google consumption / 10869 million gallons (as reported) | 2025-01-01/2025-12-31 / Global | Global operations / consumption / company_reported / High | Google / GOOGLE26 PDF p.96 Water use by data centers and offices, 2025; water methodology pp.87–88 / Not AI-only. Reported water excludes seawater and brackish water. Consumption = withdrawal minus discharge; some values may be estimated. Years and fleets differ. |
| Google withdrawal / 13562 million gallons (as reported) | 2025-01-01/2025-12-31 / Global | Operational data centers: owned and fully leased / withdrawal / company_reported / High | Google / GOOGLE26 PDF p.96 Water use by data centers and offices, 2025; water methodology pp.87–88 / Not AI-only. Reported water excludes seawater and brackish water. Consumption = withdrawal minus discharge; some values may be estimated. Years and fleets differ. |
| Google consumption / 10523 million gallons (as reported) | 2025-01-01/2025-12-31 / Global | Operational data centers: owned and fully leased / consumption / company_reported / High | Google / GOOGLE26 PDF p.96 Water use by data centers and offices, 2025; water methodology pp.87–88 / Not AI-only. Reported water excludes seawater and brackish water. Consumption = withdrawal minus discharge; some values may be estimated. Years and fleets differ. |
| Google PUE / 1.09 facility kWh / IT kWh | 2025-01-01/2025-12-31 / Global | Data-center fleet / not_applicable / company_reported / High | Google / GOOGLE26 PDF p.15, fleet-wide average PUE / Not AI-only; reporting periods, fleets and water definitions differ. |
| Microsoft withdrawal / 13266 ML | FY2025 ended 2025-06-30 / Global | Enterprise offices, owned/leased datacenters and labs / withdrawal / company_reported / High | Microsoft / MS26DATA PDF p.7, Table 8, FY25 / Not AI-only; reporting periods, fleets and water definitions differ. |
| Microsoft consumption / 8170 ML | FY2025 ended 2025-06-30 / Global | Enterprise offices, owned/leased datacenters and labs / consumption / company_reported / High | Microsoft / MS26DATA PDF p.7, Table 8, FY25 / Not AI-only; reporting periods, fleets and water definitions differ. |
| Microsoft PUE / 1.17 facility kWh / IT kWh | FY2025 ended 2025-06-30 / Global | Data-center global average / not_applicable / company_reported / High | Microsoft / MS26REPORT PDF p.23, Datacenters energy / Not AI-only; reporting periods, fleets and water definitions differ. |
| Microsoft WUE / 0.27 L/kWh (publisher definition) | FY2025 ended 2025-06-30 / Global | Owned datacenters / water_use_unspecified / company_reported / High | Microsoft / MS26REPORT PDF p.24; definition endnote 12 p.64 / Inspected 2026 definition says cooling/humidification water use divided by energy, without explicitly resolving withdrawal vs consumption and IT denominator. Not accepted as consumptive WUE by the calculator. |
| Meta withdrawal / 5637 ML | 2024-01-01/2024-12-31 / Global | Operations excluding separately disclosed construction / withdrawal / company_reported / High | Meta / META25 PDF p.9, section 3.1, 2024 column / Not AI-only; reporting periods, fleets and water definitions differ. |
| Meta consumption / 3123 ML | 2024-01-01/2024-12-31 / Global | Operations / consumption / company_reported / High | Meta / META25 PDF p.11, section 3.2, 2024 column / Not AI-only; reporting periods, fleets and water definitions differ. |
| Meta withdrawal / 4145 ML | 2024-01-01/2024-12-31 / Global | Data centers / withdrawal / company_reported / High | Meta / META25 PDF p.9, section 3.1, 2024 column / Not AI-only; reporting periods, fleets and water definitions differ. |
| Meta consumption / 2974 ML | 2024-01-01/2024-12-31 / Global | Data centers / consumption / company_reported / High | Meta / META25 PDF p.11, section 3.2, 2024 column / Not AI-only; reporting periods, fleets and water definitions differ. |
| Meta PUE / 1.08 facility kWh / IT kWh | 2024-01-01/2024-12-31 / Global | Data centers / not_applicable / company_reported / High | Meta / META25 PDF p.8, section 2.4, 2024 column / Not AI-only; reporting periods, fleets and water definitions differ. |
Sources
| Document | Scope / exact locator | Licence |
|---|---|---|
| International Energy Agency / IEA26 Key Questions on Energy and AI (2026) | All data centres; global electricity / Executive summary: updated data centre electricity demand outlook | CC BY 4.0 stated by publisher |
| Ministry of Jal Shakti / PIB / CGWB25 Findings of Dynamic Groundwater Resources Assessment | India state/UT annual groundwater; all uses / Annexure I: State-wise ground water resources of India, 2025 | Source facts attributed; no blanket licence asserted; raw document excluded |
| Elsworth et al. / Google / GOOGLE25 Measuring the environmental impact of delivering AI at Google Scale | Median Gemini Apps text prompt; direct consumptive cooling water / Section 3.3 and Table 2; May 2025 | CC BY 4.0 on article |
| World Resources Institute / WRI23 Aqueduct 4.0 technical note | Modeled water context, not local facility impact / Technical note: screening limits and future periods | Creative Commons stated; dataset-specific terms need review before ingestion |
| Microsoft / MS24 Next-generation datacenters consume zero water for cooling | Company design disclosure; not fleet conversion evidence / Mitigating energy impacts; footnote FY2024 withdrawal WUE | Publisher copyright; raw article excluded |
| Lawrence Berkeley National Laboratory / LBNL26 United States Data Center Energy Usage Report: 2025 Update | All US data centres; electricity scenarios / Publication abstract; June 2026; Reference Case | Metadata/facts attributed; full report terms need review |
| Google / GOOGLE26 Google 2026 Environmental Report | Company-reported environmental disclosure; not AI-only / PDF pp.96–97; water methodology pp.87–88; PUE p.15 | Publisher copyright; attributed extracted facts only; raw document excluded |
| Microsoft / MS26DATA Microsoft 2026 Environmental Data Fact Sheet | Company-reported environmental disclosure; not AI-only / Tables 8 and 15, PDF pp.7,25-26; inventory methods pp.18-19 | Publisher copyright; attributed extracted facts only; raw document excluded |
| Microsoft / MS26REPORT Microsoft 2026 Environmental Sustainability Report | Company-reported environmental disclosure; not AI-only / PDF pp.23-24; WUE definition endnote 12 p.64 | Publisher copyright; attributed extracted facts only; raw document excluded |
| Meta / META25 Meta 2025 Environmental Data Index | Company-reported environmental disclosure; not AI-only / FY2024; PDF pp.8-11,18; facility withdrawal p.9 | Publisher copyright; attributed extracted facts only; raw document excluded |
| Meta / META25REPORT Meta 2025 Sustainability Report | Company-reported environmental disclosure; not AI-only / 2024 reporting; current publisher reporting page | Publisher copyright; attributed extracted facts only; raw document excluded |
| Meta / Ernst & Young / META25ASSURANCE Meta Independent Accountants Review linked from 2025 report | Company-reported environmental disclosure; not AI-only / Selected metrics, water withdrawal and consumption; not a whole-report audit | Publisher copyright; attributed extracted facts only; raw document excluded |
| World Resources Institute / WRIADMIN Aqueduct 4.0 administrative rankings workbook | Demand-weighted state/province context, not facility measurement / province_baseline sheet; indicator bws; weight Tot; 1979-2019 baseline | CC BY 4.0, confirmed in wri/Aqueduct40 README |
| World Resources Institute / WRILICENCE Aqueduct 4.0 public documentation and licence | Dataset licence and documentation / README licence paragraph | CC BY 4.0 |
| WRI India / WRIINDIA26 More Than Half of India’s Data Centres Are in Water-Stressed Regions | Published India data-centre distribution analysis; not independently recreated here / 18 May 2026, main text and linked infographic; underlying sources named | Attributed summary only; no third-party infographic copied |
| Macknick et al. / NREL / NREL12 Operational water consumption and withdrawal factors for electricity generating technologies | US electricity technologies; literature synthesis, not a current global grid factor / Abstract; DOI 10.1088/1748-9326/7/4/045802; 2012 | Publisher copyright; links and attributed facts only; no raw document distributed |
| TSMC / TSMC26 TSMC 2025 Sustainability Report now available | Semiconductor manufacturing company disclosure, not AI-only allocation / 20 August 2026; Water Positive commitment section | Publisher copyright; links and attributed facts only; no raw document distributed |
| TSMC / TSMCWATER26 Supplier wastewater governance | Semiconductor supply-chain wastewater management; self-reported programme / 7 July 2026; Identifying Risks via Three Key Indicators | Publisher copyright; links and attributed facts only; no raw document distributed |
| Meta / LimnoTech / METAVWB24 2024 Volumetric Water Benefits | Company-commissioned restoration and basin context, not a causal harm study / August 2025; PDF pp.5,9,17,25; risk method, Prineville, Los Lunas, Mesa | Publisher copyright; links and attributed facts only; no raw document distributed |
What can be done
Companies: disclose comparable boundaries, test whole-system cooling choices, reduce energy per useful result, monitor seasonality and keep replenishment separate. Governments: request supply and drought plans and monitor discharge. Communities: ask for local supply and competing-use evidence. Individuals: choose adequate tools and share sourced claims with caveats; quantified personal water savings remain a gap.
Limits
No global AI water estimate, precise facility/basin overlay, independent community causal study, chip-water allocation, validated PBIX. The supplied Excel Desktop run passed nine cases and save/reopen on 22 October 2026; see the repository current validation record for the tested workbook hash. PostgreSQL execution status is recorded separately. Regional and company evidence do not establish local damage. The historical limitations file is not silently upgraded before author confirmation.
Independent portfolio research by Purushotman Mooper.