How thirsty is AI?

All US data centers, 2023: 17.4B gal/yr. LBNL's 2024 report: 66 billion liters direct site water. 66B L / 3.785 = 17.4B gal.

Source: LBNL — 2024 United States Data Center Energy Usage Report

All US AI workloads, 2024: 4B gal/yr. This is a share-allocation estimate, not an AI-specific direct-water measurement. It allocates AI a share of total fleet-average direct water based on electricity share (~23% of 176 TWh). The implied water intensity (~0.37 L/kWh) is the fleet-average LBNL site WUE, which is lower than the ~0.95 L/kWh implied by product-specific AI-serving measurements (Altman, Google Gemini).

ChatGPT (current est.): ~97M gal/yr direct, ~580M gal/yr total. This is a modeled estimate, not an OpenAI-disclosed total. It covers current ChatGPT message serving as reflected in public per-message data — not model training, not all OpenAI API traffic, and not unusually heavy multimodal or long-running agent workflows.

Traffic. In July 2025, OpenAI said users were sending more than 2.5 billion messages per day, and a separate OpenAI research paper reported 700 million weekly active users and 18 billion messages per week. In February 2026, OpenAI said ChatGPT had more than 900 million weekly active users. Scaling July 2025 message volume by WAU growth gives roughly 3.2–3.3 billion messages per day. This is slightly conservative — OpenAI also says inquiries per user are rising.

Electricity per message. Sam Altman states an average ChatGPT query uses about 0.34 Wh. Google's Gemini Apps paper reports 0.24 Wh for a median text prompt (May 2025). Altman's number is explicitly an average ChatGPT query (including coding, data analysis, and agentic workflows), so it should sit above a median text-only prompt. Ren et al.'s GPT-3 estimate of ~4 Wh per request (~12× Altman's figure) is a heavier scenario, not a good stand-in for today's average ChatGPT message. Using ~3.25B messages/day × 0.34 Wh gives ~1.1 GWh/day, ~0.40 TWh/year, and an average continuous load of ~46 MW.

Direct on-site water. Altman's figure is 0.000085 gal (0.322 mL) per query. Google's Gemini Apps measurement is 0.26 mL for a median text prompt, with Google reporting fleet WUE of 1.15 L/kWh. LBNL says average US data center site WUE was just over 0.36 L/kWh in 2023, expected to rise as hyperscale and liquid-cooled AI servers grow. Altman's own water-energy pair implies ~0.95 L/kWh, much closer to Google's AI-serving value than generic fleet averages. Because Altman's 0.322 mL is also close to Google's 0.26 mL, the most defensible center is about 0.30–0.32 mL per average message. Using 0.31 mL × 3.25B messages/day gives ~266,000 gal/day or ~97M gal/yr.

Electricity-related water. LBNL's estimate for US data-center locations is 4.52 L/kWh of indirect water consumption in 2023, based on balancing-authority grid mixes. However, as Brian Potter argues in Construction Physics, LBNL's factor is inflated because it includes water evaporated from hydroelectric reservoirs — and NREL estimates hydro plants "consume" about 18.27 gal/kWh via evaporation, roughly 40× thermoelectric consumption. That reservoir evaporation would occur whether or not a data center existed; dams serve water storage purposes independent of power generation. Excluding hydro evaporation and using a thermoelectric-only factor of ~1.8 L/kWh (NREL) is more defensible. At 0.34 Wh/query × 1.8 L/kWh this gives ~0.61 mL/query of electricity-related water, or ~525K gal/day, ~192M gal/yr. The LBNL-inclusive convention (4.52 L/kWh) would give ~482M gal/yr instead — I use the thermoelectric-only number as the default because attributing reservoir evaporation to individual electricity consumers overstates the actionable water impact.

Sources: OpenAI (2025), Sam Altman — The Gentle Singularity, Google Cloud — Measuring the Environmental Impact of AI Inference, LBNL 2024 Report

Microsoft operations, FY24: 1.53B gal/yr direct, ~15.5B with electricity. Microsoft's official FY24 environmental data fact sheet reports 5,807 megaliters total water consumption (updated estimation approach; prior years not restated). 5,807 ML / 3.785 = 1.53B gal. Global facilities and datacenters. Electricity: ~29.8 TWh from fact sheet × 0.47 gal/kWh (thermoelectric) = ~14.0B gal. Predominantly municipal water, though some facilities use recycled/reclaimed water.

Source: Microsoft FY24 Environmental Data Fact Sheet

Google operations, 2024: 7.03B gal/yr direct (freshwater consumption), ~21.5B with electricity. Google reports 4.5B gal replenished = 64% of freshwater consumption, implying 4.5/0.64 = 7.03B total. Note: this is specifically freshwater consumption, which may not be identical in boundary to Microsoft's total operational water consumption. Global operations. Electricity: Google's 2025 environmental report puts 2024 data-center electricity at ~30.8 TWh (27% year-over-year increase) × 0.47 gal/kWh = ~14.5B gal. Predominantly municipal water, though Google is expanding use of recycled and non-potable sources at some sites.

Source: Google — Sustainability Operations, Google 2025 Environmental Report

These are modeled estimates based on the best public data available, not official 2030 forecasts. All figures are direct on-site water consumption at the data center or serving layer. They do not include the much larger attributed water used to generate electricity, because that broader footprint is far more sensitive to assumptions about the 2030 grid.

All US data centers (~80B gal/yr, range 50–110B). DOE/LBNL's 2024 report puts 2023 US data center electricity at 176 TWh with 17.4B gal direct site water, and projects 325–580 TWh by 2028 with average site WUE rising to ~0.45–0.48 L/kWh. EPRI's Feb 2026 outlook extends US data center electricity to ~380–790 TWh by 2030. The IEA's 2025 base case implies ~425 TWh US by decade end. The 80B central pairs ~500 TWh (above IEA base case) with ~0.61 L/kWh implied WUE (above LBNL's near-term projected range of 0.45–0.48). A March 2026 preprint extending these frameworks lands at ~47.3B gal (optimistic efficiency), ~73.8B gal (LBNL reference), and ~89.0B gal (no improvement baseline), with a headline band of ~60–110B gal consumed annually.

US AI workloads (~45B gal/yr, range 25–70B). No official 2030 water number exists for US AI alone. In 2023, LBNL says GPU-accelerated AI servers used >40 TWh of 176 TWh total. EPRI says AI currently accounts for ~15–25% of data center electricity. By 2028, LBNL's equipment mix suggests AI-specialized servers at roughly half or more of total. The IEA says electricity demand from accelerated servers grows ~30%/yr globally to 2030. A 55–60% AI share of US data center direct water by 2030 is the central assumption (50–65% broader band), applied to the total US data center estimate.

ChatGPT (~250M gal/yr, range 70M–500M). This is global product demand mapped onto serving infrastructure, not US-only. OpenAI's 2025 usage paper reports >700M weekly active users, >2.5B messages/day by July 2025, with 5x message growth from July 2024–2025. Altman reports ~0.34 Wh and ~0.000085 gal per query. Google's 2025 Gemini Apps paper gives ~0.24 Wh and 0.26 mL per median text prompt. Central modeling: ~10B messages/day in 2030 (range 5–15B), ~0.25 mL direct water per message (range 0.15–0.35 mL), for average text-dominant interactions.

The biggest uncertainty for all US data centers is water usage effectiveness (WUE). LBNL expects average WUE to rise as hyperscale and liquid-cooled AI facilities grow, but Microsoft is deploying zero-water cooling designs for some next-gen AI data centers, and Google says water use in high-stress areas should trend lower as older assets retire. For ChatGPT, the main swing factor is whether average 2030 usage still looks like text assistance or shifts into more compute-intensive agentic and multimodal work. All figures are water consumption, not withdrawals — withdrawals would be higher (~80–150B gal/yr for all US data centers in 2030).

Sources: LBNL 2024, EPRI Feb 2026 FAQ, IEA Electricity 2025, OpenAI 2025

All electricity-related water estimates use the thermoelectric-only factor of 0.47 gal/kWh (NREL), which excludes water evaporated from hydroelectric reservoirs. As Brian Potter argues, LBNL's higher factor (1.19 gal/kWh) is inflated by hydroelectric reservoir evaporation (~18.27 gal/kWh for hydro alone), which would occur regardless of data center electricity demand.

Microsoft & Google: These are global operations. Microsoft: ~29.8 TWh from FY24 Environmental Data Fact Sheet (updated estimation approach; prior years not restated). Google: ~30.8 TWh data-center electricity from 2025 Environmental Report. Note that Microsoft reports total operational water consumption while Google reports freshwater consumption; these scopes may not be identical.

2030 estimates: ChatGPT 2030 electricity assumes ~10B messages/day × 0.34 Wh per message = ~1.24 TWh/yr. US AI workloads estimated at ~57.5% of ~500 TWh total US DC electricity (based on EPRI 380–790 TWh range and IEA's ~425 TWh base case, with central ~500 TWh).

Sources: LBNL 2024 Report, NREL, Construction Physics

Industry figures are modeled estimates derived from withdrawal data × estimated consumption ratios. The consumption ratios do most of the work and carry significant uncertainty. Auto manufacturing and steel production were removed due to obsolete or unreliable source data.

Food & beverage has wide uncertainty (~100–275B) and is shown in potable mode because food processing requires potable-quality water, though the actual municipal vs. self-supplied share is uncertain. Semiconductor public-supply share is predominantly municipal but exact percentage is unknown.

Sources: Construction Physics, Argonne GREET

These sectors use unusually large amounts of treated/public-supply potable water because food-contact, pharmaceutical, or sanitation regulations require it. The best public federal numbers for these sectors are usually wastewater-generated or process-water-discharge totals, used here as a proxy for freshwater handled. This means these figures are closer to water throughput than to strict consumptive use. Public-supply potable fractions are modeled from regulatory requirements, facility siting patterns, and federal guidance on typical water sources — not directly reported.

Meat & poultry processing: 179B gal/yr. EPA's 2023 meat and poultry products technical development document provides national production totals and average wastewater generation rates per unit of production for four segments: meat first processing (149B lb × 0.319 gal/lb = 47.6B), meat further processing (51.4B lb × 0.970 gal/lb = 49.8B), poultry first processing (105.9B lb × 0.645 gal/lb = 68.3B), poultry further processing (13.0B lb × 1.05 gal/lb = 13.7B). Total = 179B. FSIS requires that all food-contact water be potable. EPA says most facilities use drinking-water sources (public supplies or well water). Public-supply share ~60% central because slaughter/primary processing is often in rural areas with private wells.

Pharmaceutical manufacturing: 97.2B gal/yr. EPA's 1998 pharmaceutical manufacturing development document reports 297 facilities generating 266.42 MGD total wastewater (113.78 MGD noncontact cooling + 24.03 MGD ancillary + 9.29 MGD sanitary + 9.48 MGD other + remainder). 266.42 × 365 = 97.2B gal/yr. FDA states potable water is obtained "primarily from municipal water systems" and large volumes of purified water and WFI must be generated. Public-supply share ~85% central. EPA source is older but remains the best public national baseline.

Dairy processing: 45B gal/yr. USDA 2024: 225B lb milk marketed × EPA benchmark 200 gal/1000 lb milk equivalent = 45B gal/yr. The Pasteurized Milk Ordinance requires safe/sanitary water. EPA says raw water can come from wells or municipal systems. This is a central BOTEC — real plants vary widely by product mix, CIP design, and water reuse. Public-supply share ~70% central.

Industrial launderers: ~35B gal/yr. Census 2023 County Business Patterns reports 1,421 employer establishments (NAICS 812332). EPA's commercial water study benchmarks 64,090 gal/day average per establishment, with 84% going to cleaning/sanitation. 1,421 × 64,090 × 365 = 33.2B gal/yr. This row models launderers as almost entirely public-supply users because they are commercial facilities rather than self-supplied heavy industry. Total freshwater ~33–37B with slight cushion for any self-supplied sites.

Breweries: 32B gal/yr (modeled). Brewers Association 2024: 23.1M barrels craft at 13.3% market share → ~173.7M total barrels → ~5.4B gal beer. EPA/NWRAP: 4–12 gal water per gal beer; central estimate uses 6. 5.4B × 6 = 32B gal/yr (range 22–65B). No authoritative national source-mix split exists; public-supply ~75% central is a rough estimate. This row is clearly a BOTEC.

Bottled water manufacturing: 22.8B gal/yr (total only). 2024 consumption: 16.4B gal product × 1.39 gal water used per gal bottled (IBWA benchmarking study) = 22.8B gal/yr total. FDA/GAO: sources include wells, springs, and public drinking-water systems. No defensible national public-supply split is available, so no potable figure is shown.

These rows are less uniform than the data-center rows. For meat, dairy, and pharmaceuticals, the federal numbers are wastewater-generated or process-water-discharge totals used as freshwater-handled proxies. Public-supply potable fractions are modeled, not disclosed. Bottled water has no potable split. Breweries are a BOTEC from trade data + EPA process coefficients.

Crop water shows estimated consumptive use — the portion of applied irrigation water that is actually consumed via evapotranspiration (ET). Applied irrigation always exceeds consumptive use because of runoff, deep percolation, and system inefficiency.

Consumptive use ratios: Corn is predominantly center-pivot irrigated in the Great Plains with application efficiencies of 70–85% (Nebraska Extension G2345); using 0.75 as a national average. Alfalfa is disproportionately flood/surface irrigated in the West with efficiencies of 60–75%; using 0.70. These ratios are consistent with Kukal et al. (2024) and the USDA NRCS irrigation efficiency guidance.

State rows are illustrative modeled breakouts proportioned from the national consumptive total by each state's share of national applied irrigation. They should not be treated as state-validated estimates.

Sources: USDA NASS 2023 IWMS Tables 38 & 39, Kukal et al. 2024 — Consumptive Water Use of Irrigated U.S. Corn, Nebraska Extension G2345

Golf courses (2024): 531B gal/yr. GCSAA Phase 4 reports 1.63 million acre-feet total. Water sources: ~45% wells, ~25% surface, ~21% reclaimed, ~9% municipal. Arizona uses Southwest regional rate (3.3 ac-ft/ac vs. 1.1 national median).

Source: GCSAA Phase 4 Water Report (2024)

Residential lawns (est.): ~3T gal/yr, ~1.65T from public supply. EPA reports landscape irrigation at ~9B gal/day nationally. Municipal lawn portion: 13.1T × 0.60 × 0.30 × 0.70 = 1.65T. State figures are illustrative modeled breakouts, not state-validated estimates.

Sources: EPA WaterSense, USGS PP 1894D

City figures show estimated consumptive use — water evaporated, incorporated, or otherwise not returned to source. Most delivered water returns via sewer; the consumed fraction depends heavily on climate and outdoor use.

NYC and Chicago estimates use consumption ratios validated by USGS field studies. LA and Phoenix estimates infer consumption from outdoor-use fractions (nearly all outdoor use is consumed; ~80-90% of indoor use returns via sewer). Las Vegas is the only city that directly reports consumptive use. Climate is the dominant factor: cool/wet cities consume ~10-15% of deliveries; hot/dry cities consume 35-50%.

US national totals (consumptive estimates): Both national figures are estimated consumptive use, not raw withdrawals, for consistency with the rest of this visualization.

Total US freshwater consumptive use: ~30T gal/yr. Derived by applying USGS Circular 1200 (1995) sector-level consumptive fractions — the last USGS report to publish these — to USGS Circular 1441 (2015) withdrawal data: irrigation 118 Bgal/day × 58% = ~68 Bgal/day; thermoelectric 133 Bgal/day × 1.7% = ~2.3 Bgal/day; public supply 39 Bgal/day × 19% = ~7.4 Bgal/day; industrial 14.8 Bgal/day × 16% = ~2.4 Bgal/day; other ~3.2 Bgal/day. Total ~83 Bgal/day = ~30T gal/yr (~30% of 102.6T withdrawn).

Total US public water supply consumptive use: ~2.6T gal/yr. Estimated as ~20% of 13.1T gal/yr total public supply withdrawals (USGS 2020). The 20% consumptive fraction is from the same USGS 1995 methodology (~7.5 Bgal/day consumed out of ~40.2 Bgal/day withdrawn = ~19%) and is consistent with the city-level estimates on this page (12% for wet cities like NYC and Chicago, 35–50% for arid cities like LA and Phoenix).

Sources: USGS Circular 1441 (2015), USGS Professional Paper 1894D (2020)

Every category now has a + elec option on items where electricity consumption data is available. This adds the water consumed at power plants to generate the electricity used by that sector, using the thermoelectric-only factor of 0.47 gal/kWh (NREL, excluding hydro reservoir evaporation). For data centers, electricity water is a major share of the total; for farms and cities, it is negligible (<0.1%).

Industry electricity from EIA MECS 2018 (net electricity = grid purchases, excludes on-site combustion generation) for manufacturing sectors: semis (NAICS 334413), oil (324110), food & bev (311+312), paper (322), meat & poultry (3116), pharma (3254), dairy (3115). Paper & pulp generates ~66-75% of its electricity from biomass; 51 TWh is net grid only. Industrial launderers (~2.5 TWh), breweries (~3.4 TWh), and bottled water (~1.5 TWh) are rough estimates — MECS does not break out below 4-digit NAICS for beverages, and does not cover service industries. Farm pumping (~22 TWh total) proportioned by acreage share. City electricity uses rough system-energy assumptions: NYC/Chicago ~1.5 kWh/1000 gal; LA/Phoenix/Vegas ~5 kWh/1000 gal. Public water supply uses EPA WaterSense's 2.069 kWh per 1,000 gallons of drinking water.

• This visualization mixes incompatible metrics. Data center figures are on-site consumptive use. City figures are estimated consumptive use (delivered minus sewer returns). Farm figures are estimated consumptive use (applied irrigation × efficiency ratio). National totals are estimated consumptive use (USGS 1995 sector-level fractions applied to newer withdrawal data). All figures aim to show consumptive use, but the estimation methods differ across categories.
• The ChatGPT estimate is modeled, not disclosed. It is built from OpenAI's per-query energy and water figures, scaled by traffic growth implied by WAU data, and cross-checked against Google's first-party Gemini Apps measurements. The direct on-site number (~97M gal/yr) is the more defensible piece. The electricity-related piece uses a thermoelectric-only factor (~192M gal/yr) because, as Construction Physics argues, including hydroelectric reservoir evaporation overstates the actionable water impact. Under the LBNL convention (which includes hydro), the electricity piece would be ~482M gal/yr instead.
• Microsoft and Google are global operational totals, not U.S.-only or datacenter-only. Microsoft reports total operational water consumption; Google reports freshwater consumption — these scopes may not be identical. Their +electricity uses the thermoelectric factor (0.47 gal/kWh) as a rough global approximation.
• All +elec estimates use the thermoelectric-only factor (0.47 gal/kWh), excluding water evaporated from hydroelectric reservoirs. As Construction Physics argues, that reservoir evaporation would occur regardless of data center demand. LBNL's higher factor (1.19 gal/kWh, which includes hydro) would roughly 2.5× the electricity-water component.
• Public-supply potable fractions are uncertain. For data centers, semiconductors, and food processing, the true municipal share is unknown and changing. Food & beverage is shown in potable mode because it requires potable-quality water, but the actual municipal vs. self-supplied share is uncertain.
• Crop figures are estimated consumptive use, derived from USDA applied irrigation rates × consumptive use ratios (0.75 for corn, 0.70 for alfalfa). The ratios are based on typical irrigation system efficiencies and are national averages — actual consumption varies by region, soil, and method.
• Year mixing is unavoidable. 2023 LBNL, 2023 USDA, 2024 corporate, 2025 AI usage. Every figure is year-stamped.