AI Data Center Water Usage Statistics 2026
By Axis Intelligence Research
Co-authors: Sarah Mitchell, AI & Machine Learning Analyst · Aidan Jad, EV, Green Tech & Infrastructure | Last updated: October 21, 2026 | License: CC BY 4.0
Google consumed 10.9 billion gallons of water in 2025, 2.4 times its 2021 level, and Microsoft 2.16 billion. Meta consumed 0.83 billion gallons in 2024, while Amazon reports only a 2.5-billion-gallon withdrawal. In its first Hyperscaler Water Disclosure Index (HWDI), Axis Intelligence Research scores Meta 90, Microsoft 80, Google 70 and Amazon 50.
Quick Answer
U.S. data centers consumed 66 billion liters (about 17.4 billion gallons) of water on site in 2023, and nearly 800 billion liters indirectly through power generation, per Lawrence Berkeley National Laboratory. By Axis Intelligence Research’s calculation, 92.4% of that footprint sits at power plants. Among hyperscalers, Google reported the largest 2025 volume: 10.9 billion gallons consumed.
Key Findings
- Axis Intelligence Research finds that indirect water from electricity generation made up 92.4% of U.S. data centers’ combined water consumption in 2023 (800 of 866 billion liters, from LBNL’s 2024 report).
- The Axis Intelligence Research Hyperscaler Water Disclosure Index (HWDI) scores Meta 90, Microsoft 80, Google 70 and Amazon 50 out of 100 as of October 21, 2026, based on eight weighted disclosure criteria.
- Axis Intelligence Research calculates that Google consumed 0.944 liters of water per kWh of electricity in 2025, 4.3 times Microsoft’s 0.221 L/kWh in fiscal 2025, using each company’s reported total water consumption and total electricity.
- Axis Intelligence Research calculates that Microsoft’s Phoenix data centers withdrew 74.7 times more water per kilowatt-hour than its Dublin campus in fiscal 2025 (1.028 vs 0.014 liters per kWh of site electricity).
- Meta’s purchased electricity carried 72,207 megaliters of embedded water in 2024, 24.3 times the 2,974 megaliters its data centers consumed directly, per Axis Intelligence Research analysis of Meta’s 2025 Environmental Data Index.
How Much Water Do AI Data Centers Use in the U.S.?
U.S. data centers consumed 66 billion liters (about 17.4 billion gallons) of water on site in 2023, mostly in cooling towers and humidification. That is 3.1 times the 21.2 billion liters consumed in 2014. The water used to generate their electricity was nearly 800 billion liters (about 211 billion gallons) the same year. All three figures come from the 2024 United States Data Center Energy Usage Report, which Lawrence Berkeley National Laboratory (LBNL) produced for the Department of Energy (pp. 55–57).
| Metric (U.S., all data centers) | 2023 | 2028 outlook | Source |
|---|---|---|---|
| Electricity use | 176 TWh | 325–580 TWh | LBNL, Dec. 2024 |
| Direct (on-site) water consumption | 66 billion L (≈17.4 billion gal) | Hyperscale alone: 60–124 billion L | LBNL, Dec. 2024 |
| Hyperscale + colocation share of direct water | 84% | — | LBNL, Dec. 2024 |
| Indirect water (electricity generation) | nearly 800 billion L (≈211 billion gal) | — | LBNL, Dec. 2024 |
| Indirect water per kWh used by data centers | 4.52 L/kWh (U.S. grid average: 4.35) | — | LBNL, Dec. 2024 |
| Average site Water Usage Effectiveness (WUE) | just over 0.36 L/kWh | 0.45–0.48 L/kWh | LBNL, Dec. 2024 |
| Indirect share of combined water | 92.4% | — | Axis Intelligence Research calculation |
| Indirect-to-direct multiple | 12.1× | — | Axis Intelligence Research calculation |
Gallon figures are Axis Intelligence Research conversions; LBNL reports liters.
Water Usage Effectiveness (WUE) is the liters of water a data center uses per kilowatt-hour of IT equipment electricity. LBNL measures it on a consumption basis: water that evaporates and does not return to the local source.
The 92.4% that never shows up in a cooling-tower debate
Axis Intelligence Research divided LBNL’s indirect figure by the combined total: 800 ÷ (800 + 66) = 92.4%. Put plainly, for every liter a U.S. data center evaporates on site, about 12 more are consumed at the power plants feeding it. LBNL notes that its indirect estimate uses the local balancing-authority grid mix and does not reflect individual power purchase agreements.
LBNL projects that hyperscale data centers alone will consume 60 to 124 billion liters of water on site in 2028. The top of that range is 1.88 times what all U.S. data centers consumed in 2023.
Sarah Mitchell, analyst read: The fight over evaporative cooling is real, but it is a fight over the smaller line item. LBNL expects average site WUE to rise from about 0.36 to between 0.45 and 0.48 L/kWh after 2023, as more liquid-cooled AI racks come online, while electricity demand climbs from 176 TWh toward 325–580 TWh by 2028. At 4.52 liters of grid water per kWh, every efficiency point gained in the server hall is outweighed by the generation mix on the interconnection. A gas-heavy grid in a dry basin can erase any chiller decision made at the site.
How Much Water Do Google, Microsoft, Amazon and Meta Use?
The four largest U.S. hyperscalers now all publish an annual water number. They do not publish the same number. The table below records each figure exactly as reported, with its measurement boundary.
| Company | Latest volume | Year | Metric | Boundary | Source |
|---|---|---|---|---|---|
| 10,869 million gal (≈41 billion L) | 2025 | Consumption | Data centers + offices (10,523 million gal data centers) | Google 2026 Environmental Report | |
| Microsoft | 8,170 ML (≈2.16 billion gal) | FY2025 | Consumption | Company-wide incl. data centers | Microsoft 2026 Environmental Data Fact Sheet |
| Amazon | 2.5 billion gal (9.4 billion L) | 2025 | Withdrawal | Leased, owned and shared data centers | AWS Sustainability Report 2025 summary |
| Meta | 3,123 ML (≈0.83 billion gal) | 2024 | Consumption | Company (2,974 ML data centers) | Meta 2025 Environmental Data Index |
ML = megaliters (one million liters). Conversions use 3.785411784 liters per U.S. gallon.
Google: 10.9 billion gallons, up 34%
Google consumed 10,869 million gallons across data centers and offices in 2025, a 34% increase, while its electricity use grew 37%. It withdrew 14,689 million gallons. Data centers accounted for 10,523 million gallons of consumption (96.8%), according to the data tables of the Google 2026 Environmental Report, which were subject to third-party limited assurance.
| Google, global operations | 2021 | 2022 | 2023 | 2024 | 2025 | Source |
|---|---|---|---|---|---|---|
| Water withdrawal (million gal) | 6,297 | 7,600 | 8,653 | 11,011 | 14,689 | Google 2026 Env. Report |
| Water consumption (million gal) | 4,562 | 5,565 | 6,352 | 8,135 | 10,869 | Google 2026 Env. Report |
| Freshwater consumption (million gal) | — | 4,770 | 5,601 | 7,240 | 9,947 | Google 2026 Env. Report |
| Water replenished (million gal) | — | 271 | 1,036 | 4,590 | 7,717 | Google 2026 Env. Report |
| Replenished, % of freshwater consumption | — | 6% | 18% | 63% | 78% | Google 2026 Env. Report |
Google recalculated its 2024 replenishment rate to 63% (previously reported as 64%).
Google’s consumption grew 2.38 times between 2021 and 2025. Its 165 stewardship projects in 97 watersheds replenished 7,717 million gallons in 2025, which Google reports as 78% of its freshwater consumption (9,947 million gallons). Measured against total consumption, the share is 71.0%, by Axis Intelligence Research’s calculation. Of Google’s 2025 freshwater withdrawal, 13% came from sources at high risk of depletion or scarcity, 15% from medium-risk sources and 72% from low-risk sources.
Which Google Data Centers Consume the Most Water?
Google publishes withdrawal, discharge and consumption for each data center location. Five sites account for 44.8% of its 2025 data center consumption, and Council Bluffs, Iowa, alone for 12.8%, by Axis Intelligence Research’s calculation.
| Google data center location | Withdrawal (million gal) | Consumption (million gal) | Source |
|---|---|---|---|
| Council Bluffs, IA | 1,746.4 | 1,346.0 | Google 2026 Env. Report |
| Mayes County, OK | 1,396.7 | 1,081.4 | Google 2026 Env. Report |
| Berkeley County, SC | 988.5 | 901.7 | Google 2026 Env. Report |
| New Albany, OH | 1,008.1 | 835.7 | Google 2026 Env. Report |
| Papillion, NE | 677.8 | 548.2 | Google 2026 Env. Report |
| The Dalles, OR | 654.3 | 469.0 | Google 2026 Env. Report |
| St. Ghislain, Belgium | 714.4 | 450.8 | Google 2026 Env. Report |
| Montgomery County, TN | 499.5 | 424.7 | Google 2026 Env. Report |
| Eemshaven, Netherlands | 532.2 | 405.7 | Google 2026 Env. Report |
| Douglas County, GA | 440.6 | 367.1 | Google 2026 Env. Report |
| Mesa, AZ | 19.4 | 9.7 | Google 2026 Env. Report |
| Hamina, Finland | 4.3 | 0.4 | Google 2026 Env. Report |
| Dublin, Ireland | 0.5 | 0.1 | Google 2026 Env. Report |
Potable water made up 80.7% of Google’s 2025 data center withdrawal. Douglas County, Georgia, runs mostly on reclaimed wastewater: 394.3 of its 440.6 million gallons withdrawn.
Sarah Mitchell: Google’s data center consumption runs at 0.939 liters per kWh of data center electricity (10,523 million gallons over 42.4 TWh), an Axis Intelligence Research calculation. That reflects Google’s choice of water cooling, which its report describes as requiring less energy than air cooling. The geography matters as much as the volume: Google’s largest water sites sit in the Midwest and Southeast, while its Mesa, Arizona campus consumed just 9.7 million gallons.
Microsoft: consumption up 22.1%, intensity flat
Microsoft consumed 8,170 megaliters of water in fiscal 2025, up from 6,693 ML, and withdrew 13,266 ML, per its 2026 Environmental Data Fact Sheet, reviewed by Deloitte. Consumption rose 22.1% while electricity use rose 24.1%. Axis Intelligence Research’s company-wide ratio of water consumed per kWh of electricity therefore moved only from 0.224 to 0.221 L/kWh.
Microsoft reports a fleet WUE of 0.27 L/kWh in 2025, down from 0.30 in fiscal 2024 and 2.3 for its first-generation data centers, according to a June 24, 2026 Microsoft post. About 90% of its owned fleet now uses low- or zero-water cooling. Half of its fiscal 2025 withdrawals (50%) came from water-stressed areas.
Amazon: a withdrawal number, not a consumption number
Amazon reported withdrawing about 2.5 billion gallons across its global data center footprint in 2025, its first company-wide water figure, in a June 11, 2026 post. Its AWS Sustainability Report 2025 summary gives the same total as 9.4 billion liters and does not report consumption.
AWS’s withdrawal-based WUE fell from 0.25 L/kWh in 2021 and 0.15 in 2024 to 0.12 in 2025. That is a 20% improvement in one year and 52% since 2021. Free-air cooling runs about 90% of the time, and water is used to cool incoming air only above roughly 85°F. 22% of withdrawals at leased and owned sites occurred in high or extremely high water-stress regions.
Meta: the only hyperscaler reporting its grid water
Meta consumed 3,123 ML in 2024 (2,974 ML in data centers) and withdrew 5,637 ML, per its 2025 Environmental Data Index. Its data center WUE, on a withdrawal basis, was 0.19 L/kWh. Meta’s consumption in high or extremely high water-stress areas rose 48.4%, from 504 ML to 748 ML, to reach 24.0% of its total.
In July 2026, Fortune reported that Cheyenne, Wyoming, officials traced a rare bacterium in the city’s reclaimed irrigation water to construction discharges at Meta’s data center campus. The bacterium did not reach drinking water, and Meta says testing commissioned by its general contractor found no trace of it.
Why Hyperscaler Water Numbers Can’t Be Compared Directly
Withdrawal is all the water a facility draws from a utility, river or aquifer. Consumption is the part that does not go back, mostly evaporated in cooling. A withdrawal figure is always larger than the matching consumption figure.
The gap is not small. Axis Intelligence Research computes consumption at 74.0% of withdrawal for Google (2025), 61.6% for Microsoft (FY2025) and 55.4% for Meta (2024). Setting Amazon’s 2.5-billion-gallon withdrawal beside Google’s 10.9-billion-gallon consumption, as much of the coverage of Amazon’s disclosure did, compares two different measurements.
| Company | Consumption | Withdrawal | Consumption ÷ withdrawal | WUE basis | Source |
|---|---|---|---|---|---|
| Microsoft (FY2025) | 8,170 ML | 13,266 ML | 61.6% | Fleet WUE 0.27 L/kWh | Microsoft Fact Sheet 2026 |
| Meta (2024) | 3,123 ML | 5,637 ML | 55.4% | Withdrawal, 0.19 L/kWh | Meta Environmental Data Index 2025 |
| Amazon (2025) | Not reported | 9.4 billion L | — | Withdrawal, 0.12 L/kWh | AWS Sustainability Report 2025 |
| Google (2025) | 10,869 million gal | 14,689 million gal | 74.0% | No fleet WUE published | Google Environmental Report 2026 |
Water consumed per kWh: Google vs Microsoft vs Meta
WUE values cannot be compared across these companies. Axis Intelligence Research therefore divided each company’s reported water consumption by its reported electricity use. These are not WUE values, because the denominators include non-IT load.
| Company | Water consumed | Electricity | Liters per kWh | Source |
|---|---|---|---|---|
| Google, company (2025) | 10,869 million gal | 43,586,600 MWh | 0.944 | Google 2026 Env. Report |
| Google, company (2024) | 8,135 million gal | 31,713,900 MWh | 0.971 | Google 2026 Env. Report |
| Google, data centers (2025) | 10,523 million gal | 42,415,800 MWh | 0.939 | Google 2026 Env. Report |
| Microsoft, company (FY2025) | 8,170 ML | 37,026,353 MWh | 0.221 | Microsoft Fact Sheet 2026 |
| Microsoft, company (FY2024) | 6,693 ML | 29,829,540 MWh | 0.224 | Microsoft Fact Sheet 2026 |
| Meta, data centers (2024) | 2,974 ML | 18,061,781 MWh | 0.165 | Meta Environmental Data Index 2025 |
On this measure, Google consumed 4.3 times as much water per kWh as Microsoft. Both companies held their ratios roughly flat year over year, which means their water growth tracks their electricity growth.
Inside Amazon’s “7x” water efficiency claim
Amazon compares its 0.12 L/kWh with an industry average of 0.84 L/kWh. Its own water stewardship page explains where that 0.84 comes from: a 0.56 L/kWh consumption estimate from a 2025 study, multiplied by an assumed three cycles of concentration to approximate withdrawal.
So the ratio sets a measured withdrawal figure against an estimated one. LBNL’s national site average, a consumption measure, is just over 0.36 L/kWh.
Sarah Mitchell: Amazon’s improvement is real: the drop from 0.25 to 0.12 L/kWh is on a consistent withdrawal basis. The “7x” headline is a modeling choice. Procurement teams comparing cloud regions should ask every provider the same two questions: withdrawal or consumption, and over which sites.
Hyperscaler Water Disclosure Index (HWDI) 2026: Who Discloses the Most?
The Hyperscaler Water Disclosure Index (HWDI) is an Axis Intelligence Research metric. It scores how completely each company’s public documents let an outsider measure its water footprint, on a 0–100 scale. It does not score how much water a company uses. This is the baseline reading, dated October 21, 2026.
| Criterion | Weight | Microsoft | Meta | Amazon | |
|---|---|---|---|---|---|
| C1 Absolute annual water volume published | 10 | ✓ | ✓ | ✓ | ✓ |
| C2 Data for 2025 (calendar or fiscal) available | 10 | ✓ | ✓ | — (latest: 2024) | ✓ |
| C3 Consumption reported, not only withdrawal | 15 | ✓ | ✓ | ✓ | — |
| C4 Facility-level water data | 15 | ✓ | ✓ | ✓ | — |
| C5 Fleet WUE published | 10 | — | ✓ | ✓ | ✓ |
| C6 Water embedded in purchased electricity | 20 | — | — | ✓ | — |
| C7 Water-stress breakdown | 10 | ✓ | ✓ | ✓ | ✓ |
| C8 Replenishment volume published | 10 | ✓ | ✓ | ✓ | ✓ |
| HWDI score (out of 100) | 100 | 70 | 80 | 90 | 50 |
Sources: Google 2026 Environmental Report and operations page; Microsoft 2026 Environmental Data Fact Sheet; Meta 2025 Environmental Data Index; AWS Sustainability Report 2025 summary and Amazon’s June 2026 post. Formula: HWDI = Σ (weight × criterion met). Component scores are in the CSV.
What separates the scores
- Grid water (C6) carries the highest weight because LBNL’s data places 92.4% of the U.S. footprint there. Only Meta publishes it. Amazon told Latitude Media in August 2026 that it tracks the figure but does not publish it.
- Amazon loses 30 points on C3 and C4: it reports a global withdrawal total and a stress breakdown, but no consumption and no site-level figures.
- Google publishes five years of withdrawal, discharge and consumption, a 40-location site table and third-party limited assurance, but no fleet WUE and no grid-water figure.
- Microsoft is the only company of the four whose site table pairs water withdrawal with electricity, which makes the regional analysis below possible.
Sarah Mitchell: A 90 for Meta is a disclosure score, not an environmental score. Meta also shows the steepest growth in grid water of any company here: +126% from 2021 to 2024. Transparency is what lets you see that.
Where Do Microsoft’s Data Centers Use the Most Water per kWh?
Microsoft’s fiscal 2025 fact sheet lists electricity and water withdrawal for each owned data center location. Dividing the two produces a site-level withdrawal intensity that Microsoft does not publish. This is an Axis Intelligence Research calculation, and it is not a WUE because the denominator is total site electricity rather than IT load.
| Location | Site electricity (MWh) | Withdrawal (ML) | Withdrawal intensity (L/kWh) | Replenishment ÷ withdrawal | Source |
|---|---|---|---|---|---|
| Singapore | 338,845 | 423 | 1.248 | — (99% non-potable) | Microsoft Fact Sheet 2026, Table 15 |
| Phoenix (AZ) | 954,206 | 981 | 1.028 | 2.73 | Microsoft Fact Sheet 2026, Table 15 |
| Quincy (WA) | 1,381,569 | 1,292 | 0.935 | 0.64 | Microsoft Fact Sheet 2026, Table 15 |
| Chicago (IL) | 369,328 | 204 | 0.552 | — | Microsoft Fact Sheet 2026, Table 15 |
| San Antonio (TX) | 1,440,710 | 420 | 0.292 | 2.20 | Microsoft Fact Sheet 2026, Table 15 |
| Ashburn (VA) | 828,925 | 239 | 0.288 | — | Microsoft Fact Sheet 2026, Table 15 |
| Cheyenne (WY) | 1,091,460 | 188 | 0.172 | 0.32 | Microsoft Fact Sheet 2026, Table 15 |
| East Wenatchee (WA) | 299,727 | 51 | 0.170 | — | Microsoft Fact Sheet 2026, Table 15 |
| Des Moines (IA) | 2,152,335 | 264 | 0.123 | 0.46 | Microsoft Fact Sheet 2026, Table 15 |
| Boydton (VA) | 3,113,847 | 362 | 0.116 | — | Microsoft Fact Sheet 2026, Table 15 |
| Hollands Kroon (NL) | 1,291,170 | 46 | 0.036 | — | Microsoft Fact Sheet 2026, Table 15 |
| Gävle–Sandviken (SE) | 704,549 | 14 | 0.020 | — | Microsoft Fact Sheet 2026, Table 15 |
| Dublin (IE) | 1,308,581 | 18 | 0.014 | — | Microsoft Fact Sheet 2026, Table 15 |
Replenishment is shown only where Microsoft reports a delivered volume for the location.
Climate is the main driver. Microsoft says water is needed for cooling as much as 40% of the year in Phoenix, about 10% in Virginia, and under 5% in Dublin and Amsterdam. Boydton, Virginia, draws the most electricity of any Microsoft location in the table (3.1 TWh) at one of the lowest intensities, 0.116 L/kWh.
Sarah Mitchell: Siting decides this more than any other choice. Microsoft’s Dublin and Phoenix campuses use broadly similar cooling philosophies, yet Phoenix withdraws 74.7 times more water per kWh. Replenishment is where it should be in Phoenix, at 2.73 times withdrawal. In Cheyenne and Des Moines, delivered replenishment still covers less than half of the water withdrawn.
What Is the Hidden Water Footprint of AI Data Centers?
The hidden footprint is the water consumed by thermoelectric and other power plants to produce a data center’s electricity. It never passes through the facility’s meter.
| Measure | Value | Year | Source |
|---|---|---|---|
| U.S. data centers, indirect water | nearly 800 billion L | 2023 | LBNL 2024 |
| U.S. indirect-to-direct multiple | 12.1× | 2023 | Axis Intelligence Research calculation |
| Meta, water embedded in purchased electricity | 72,207 ML | 2024 | Meta Environmental Data Index 2025 |
| Meta, embedded ÷ data center consumption | 24.3× | 2024 | Axis Intelligence Research calculation |
| Meta, embedded water change 2021–2024 | +126.2% | 2024 | Axis Intelligence Research calculation |
Meta’s multiple (24.3×) is about double LBNL’s national average (12.1×). Meta’s figure comes from its own grid mix and accounting method, which the data index does not detail, so the two are not strictly comparable. Even at the national rate, a water-positive pledge that covers only on-site water addresses less than a tenth of the footprint.
How Much Water Does ChatGPT Use per Query?
OpenAI CEO Sam Altman wrote in June 2025 that an average ChatGPT query uses about 0.000085 gallons (about 0.32 milliliters) of water. OpenAI has not published the scope or method behind that figure.
The UC Riverside paper Making AI Less “Thirsty” (version 5, March 2025) estimates that GPT-3 needs a 500 mL bottle of water for roughly 10 to 50 medium-length responses. That is 10–50 mL per response, an Axis Intelligence Research division, on a scope that includes electricity-related water. The same paper estimates that training GPT-3 in Microsoft’s U.S. data centers directly evaporated 700,000 liters, or about 5.44 million liters including electricity-related water.
| Estimate | Water per response/query | Scope | Source |
|---|---|---|---|
| OpenAI (Altman), 2025 | 0.32 mL | Not disclosed | The Gentle Singularity |
| UC Riverside, GPT-3 | 10–50 mL | On-site plus electricity | arXiv:2304.03271v5 |
ChatGPT handled 18 billion messages a week from 700 million weekly users by July 2025, according to NBER Working Paper 34255. At Altman’s per-query rate, that volume works out to about 5.79 million liters a week, an Axis Intelligence Research calculation.
Sarah Mitchell: The 31-to-155-fold spread between these estimates is mostly a question of where you draw the boundary, not a measurement disagreement. A per-query figure without a stated scope cannot be audited, and that applies to the 0.32 mL as much as to the viral half-liter.
How Much Water Do Texas Data Centers Use?
Nobody knows precisely, and the State of Texas confirmed as much in June 2026. At a June 23 hearing of the Texas House Committee on Natural Resources, lawmakers heard that only 17% of the 341 data centers surveyed by the Texas Water Development Board (TWDB) answered its mandatory 2025 water-use survey, KUT reported. Two years earlier, the board surveyed 22 data centers.
| Texas metric | Value | Source |
|---|---|---|
| Data centers surveyed by TWDB, 2025 | 341 | KUT, June 26, 2026 |
| TWDB survey response rate, 2025 | 17% | KUT, June 26, 2026 |
| PUCT survey respondents | 28 companies, 92 data center and crypto-mining facilities | KUT, June 26, 2026 |
| Maximum fine, Class C misdemeanor | $500 | Texas Penal Code §12.23 |
| Current data center water use (cooling + electricity) | 25 billion gal/yr | HARC, Jan. 21, 2026 |
| 2030 projection | 29–161 billion gal/yr (up to 2.7% of state use) | HARC, Jan. 21, 2026 |
Not answering the TWDB survey is a Class C misdemeanor, which under the Penal Code carries a fine of no more than $500. The Houston Advanced Research Center (HARC) estimates current Texas data center water use at 25 billion gallons a year. Its high case for 2030, 161 billion gallons, is 6.4 times that level.
Sarah Mitchell: The high case is not a forecast of evaporation. HARC’s number includes the water behind the electricity, which is why Texas’s grid mix matters as much as its chillers. With an 83% non-response rate, the state’s 2027 water plan is drafting against estimates, not meter readings.
Are Google, Microsoft, Amazon and Meta on Track to Be Water Positive?
“Water positive” means returning more water to watersheds than an operation uses. Each company defines the denominator differently.
| Company | Target | Latest progress | Source |
|---|---|---|---|
| Replenish 120% of freshwater consumption by 2030 | 78% (2025), up from 63% (2024), 18% (2023) and 6% (2022) | Google 2026 Environmental Report | |
| Microsoft | Replenish more than it withdraws by 2030 | Reports it replenished more than it withdrew in FY2025 | Microsoft, June 24, 2026 |
| Amazon | Water positive for data centers by 2030 | 75% (2025), up from 53% (2024) | AWS Sustainability Report 2025 |
| Meta | Water positive by 2030 (stress-weighted) | 6,017 ML restored vs 3,123 ML consumed (2024) | Meta Environmental Data Index 2025 |
At its 2025 freshwater consumption level, Google would need to replenish about 11.94 billion gallons a year to reach 120%, an Axis Intelligence Research calculation (1.2 × 9,947 million). Google expects its portfolio to deliver 19,722 million gallons a year in 2030 once fully implemented, but its freshwater consumption rose 37% in 2025 alone. Microsoft has set a 40% data center water-intensity reduction target for 2030 against 2022 and reports 25% achieved.
All four commitments cover on-site or operational water. None includes the grid water that LBNL’s data shows dominates the national footprint.
UN Secretary-General António Guterres launched the AI Environmental Transparency Initiative on June 23, 2026, at London Climate Action Week. In his address, he called on “every major AI company” to disclose the carbon, water and land footprints of its systems and to power every data center with renewable energy by 2030. The initiative builds on a United Nations University report on AI’s environmental costs.
Methodology
Collection. Axis Intelligence Research read each company’s latest primary water disclosure and LBNL’s 2024 report, retrieved on October 21, 2026, and recorded every figure in the accompanying CSV. Each row lists the source organization, document, URL, retrieval date and as-of date. Press reports are used only for facts they reported first (KUT on the Texas hearing, Fortune on Cheyenne, Latitude Media on grid-water reporting), and are flagged is_primary = no.
Axis Intelligence Research calculations (flagged axis_calculated = yes in the CSV, with the formula in method_note):
- Indirect share = LBNL indirect ÷ (indirect + direct) = 800 ÷ 866 billion liters = 92.4%.
- Site withdrawal intensity (L/kWh) = withdrawal (ML) × 1,000 ÷ site electricity (MWh), from Microsoft’s Table 15. It uses total site electricity, so it is not comparable to WUE.
- Consumption ÷ withdrawal = reported consumption ÷ reported withdrawal, same year and boundary.
- Embedded multiple = Meta embedded electricity water ÷ Meta data center consumption.
- Water consumed per kWh = reported water consumption (L) ÷ reported electricity (kWh), same company and year.
- Weekly ChatGPT water = 18 billion messages × 0.000085 gal × 3.785411784 L/gal. This assumes Altman’s average applies to every message.
HWDI construction. HWDI = Σ (criterion weight × 1 if met, 0 if not), across eight criteria summing to 100 (see table above). A criterion counts as met only when the item appears in the company documents listed in the CSV source_url column. Weights favor the measures that matter most for a watershed: grid water (20), consumption basis (15) and facility-level data (15). Axis Intelligence Research will re-score when any company publishes new water data and will log each reading by date. Companies can submit documentation to [email protected].
Scope. Volumes are reported by each company on its own boundary: data centers and offices (Google), company-wide (Microsoft and Meta), or data centers only (Amazon). Axis Intelligence Research does not merge these into a single total. Meta’s latest available data covers 2024.
About This Dataset
Title: AI Data Center Water Usage Statistics 2026 (Hyperscaler Water Disclosure Index dataset) Creator and publisher: Axis Intelligence Research
Contents: 346 rows covering U.S. data center water (LBNL 2023), Google 2021–2025 including 18 data center locations, Microsoft FY2024–FY2025 including 15 data center locations, Amazon 2021–2025, Meta 2021–2024, Texas survey and projection data, ChatGPT water estimates, and HWDI component scores.
Format: CSV, UTF-8, one observation per row, with source columns (source_org, source_document, source_url, retrieved_date, is_primary, axis_calculated, method_note).
License: CC BY 4.0. Free to reuse with attribution.
Mirrors: Hugging Face, Kaggle and GitHub (links added on upload).
How to Cite This Data
APA: Axis Intelligence Research, & Mitchell, S. (2026, October 21). AI data center water usage statistics 2026: What Google, Microsoft, Amazon and Meta actually disclose. Axis Intelligence. https://axis-intelligence.com/ai-data-center-water-usage-statistics/
MLA: Axis Intelligence Research, and Sarah Mitchell. “AI Data Center Water Usage Statistics 2026: What Google, Microsoft, Amazon and Meta Actually Disclose.” Axis Intelligence, 21 Oct. 2026, axis-intelligence.com/ai-data-center-water-usage-statistics/.
Chicago: Axis Intelligence Research, and Sarah Mitchell. “AI Data Center Water Usage Statistics 2026: What Google, Microsoft, Amazon and Meta Actually Disclose.” Axis Intelligence, October 21, 2026. https://axis-intelligence.com/ai-data-center-water-usage-statistics/.
Short form for the index: Axis Intelligence Research, Hyperscaler Water Disclosure Index (HWDI), October 21, 2026 reading.
Frequently Asked Questions
Is Amazon’s 2.5 billion gallons comparable to Google’s 10.9 billion?
No. Amazon reports withdrawal for its data centers, while the 10.9 billion figure is Google’s consumption. Google’s own withdrawal was 14,689 million gallons in 2025. Consumption ran at 74.0% of withdrawal at Google, 61.6% at Microsoft and 55.4% at Meta, by Axis Intelligence Research’s calculation; Amazon has not published a consumption figure.
Why does Microsoft’s Phoenix campus use so much more water per kWh than Dublin?
Mostly because of climate. Microsoft says Phoenix sites need water for cooling up to 40% of the year, compared with under 5% in Dublin. Axis Intelligence Research measures Phoenix at 1.028 L of withdrawal per kWh of site electricity in fiscal 2025, against 0.014 in Dublin, a 74.7× gap.
What WUE is “good” for a hyperscale data center in 2026?
The answer depends on the basis. Microsoft reports 0.27 L/kWh, Meta 0.19 (withdrawal) and AWS 0.12 (withdrawal). LBNL’s U.S. average is just over 0.36 L/kWh on a consumption basis. Comparing WUE values is only meaningful when the basis and the facility set match.
Which tech company reports the water used to generate its electricity?
Among Google, Microsoft, Amazon and Meta, only Meta does: 72,207 megaliters embedded in purchased electricity in 2024, per its 2025 Environmental Data Index. LBNL estimates this indirect water at nearly 800 billion liters for all U.S. data centers in 2023.
Does one ChatGPT question really use a bottle of water?
No. UC Riverside estimates a 500 mL bottle covers 10 to 50 GPT-3 responses, including electricity-related water, which is about 10–50 mL each. Sam Altman puts an average ChatGPT query at 0.32 mL but has not disclosed the scope.
How much water do Texas data centers use?
HARC estimates about 25 billion gallons a year, including water for electricity, and 29 to 161 billion gallons by 2030. Official figures are thin: only 17% of the 341 data centers surveyed by the Texas Water Development Board responded in 2025.
Do “water positive” pledges cover a data center’s full water footprint?
No. Google, Microsoft, Amazon and Meta measure progress against on-site or operational water. LBNL’s data puts about 92.4% of U.S. data centers’ combined water consumption at power plants, outside these pledges.
How fast is AI data center water use growing?
Google’s consumption grew 34% in 2025 and Microsoft’s grew 22.1% in fiscal 2025. Meta’s embedded grid water grew 126.2% from 2021 to 2024. LBNL expects average site WUE to rise from about 0.36 to 0.45–0.48 L/kWh after 2023 as liquid-cooled AI capacity expands.
Related Axis Intelligence Research
- AI Data Center Energy Consumption Statistics: the electricity side of the same LBNL data.
- Data Center Grid Impact Statistics: where data center load is landing on the U.S. grid.
- AI Data Center Statistics: capacity, investment and power at a glance.
- Data Center Moratorium Tracker: local and state pauses, many citing water.
- AI Data Center Forecast 2030: the build-out behind the water numbers.
- Data Center Electricity Demand Tracker: monthly demand signals.