Home Writing Projects Lab CV
Water arithmetic

Drinking AI

Asking ChatGPT 30 questions a day for a year uses about a fifth of the water behind one beer — counting data-center cooling plus the water used to generate the query's electricity, the closest AI analog to how the drink figures count the water behind the crops. Pick a drink, shift the scope, or count by agentic tasks; then compare whole US drink categories with the modeled water footprint of all AI in 2025, and with a 2026 extrapolation (a Claude estimate).

Sam Altman's September 2026 line — one California almond takes the water of 38,000 queries — works out at OpenAI's 0.32 mL per query and the California-specific almond footprint: 12.2 L against 12.3 L. At this page's default scope, one almond ≈ 9,660 queries, and a glass of almond milk, about four almonds, ≈ 38,100.

One drink, or one almond, in ChatGPT queries

One beer is about 53,300 average ChatGPT queries, enough water for 30 queries a day for 5 years.

One beer
average ChatGPT queries

Enough water for 30 queries a day for 5 years.

All scopes: 2,370–331,000 queries.

Water counted per unit
Counting by
1 drop = 1,000 queries 53 drops
Aggregate view

A year of US drinks vs a year of global AI

A 2026 Patterns paper models the water footprint of all AI systems worldwide at 312.5 billion liters to 764.6 billion liters in 2025. The bars share one linear scale.

Annual water footprints on one scale: global AI 312.5 billion liters to 764.6 billion liters in 2025, projected 468.8 billion liters to 1.1 trillion liters in 2026 (extrapolation); US coffee 24.9 trillion liters (33–80× AI); US milk 19.8 trillion liters (26–63× AI); US soda 15.3 trillion liters (20–49× AI); US beer 7 trillion liters (9.2–22× AI); US wine 2.9 trillion liters (3.8–9.3× AI); US orange juice 2.8 trillion liters (3.7–9.1× AI); US tea 1.1 trillion liters (1.4–3.4× AI); US plant-based milk 402.4 billion liters (0.5–1.3× AI); US bottled water 88.4 billion liters (0.1–0.3× AI).

All global AI 2025, modeled 312.5–764.6 billion liters

Low to high estimate: 1.3–3.1% of the coffee bar. World total, all AI systems, consumption. The dashed line on each drink bar marks the high estimate.

All global AI 2026, projected extrapolation 468.8 billion–1.1 trillion liters

Claude estimate: 2025 scaled by 1.5× AI-power growth (published rates span 1.3–2.45×): 1.9–4.6% of the coffee bar.

US coffee 2024/25 24.9 trillion liters 33–80× AI

1.57 billion kg of green coffee (26.1M 60-kg bags) · 15,897 L per kg green coffee — measured by bean mass, not brewed volume · USDA FAS coffee report, Mekonnen & Hoekstra 2011

US milk 2025 19.8 trillion liters 26–63× AI

42.8 billion lb of fluid milk sold · 1,020 L per kg, same source as the calculator · USDA ERS dairy data, Mekonnen & Hoekstra 2012

US soda 2025 15.3 trillion liters 20–49× AI

12.0 billion gallons, derived from IBWA per-capita figures · 338 L per L — the low end of the published range · IBWA 2026 progress report, Ercin, Aldaya & Hoekstra 2011

US beer 2023 7 trillion liters 9.2–22× AI

6.17 billion gallons (NIAAA) · 300 L per L, same source as the calculator · NIAAA Surveillance Report 122, Mekonnen & Hoekstra 2011

US wine 2023 2.9 trillion liters 3.8–9.3× AI

0.89 billion gallons (NIAAA) · 860 L per L, same source as the calculator · NIAAA Surveillance Report 122, Mekonnen & Hoekstra 2011

US orange juice 2022 2.8 trillion liters 3.7–9.1× AI

0.71 billion gallons (USDA ERS, latest year in the series) · 1,060 L per L, same source as the calculator · USDA ERS fruit-juices table, Mekonnen & Hoekstra 2011

US tea 2024 soft estimate 1.1 trillion liters 1.4–3.4× AI

268 million lb of tea imported (Tea Association projection; 86 billion servings) · 8,856 L per kg of dry tea, by leaf mass like coffee; the softest figure on the chart · Tea Association of the USA, Mekonnen & Hoekstra 2011

US plant-based milk 2025 soft estimate 402.4 billion liters 0.5–1.3× AI

0.36 billion gallons of almond, oat, soy and other drinks (Circana retail data via NMPF) · 297 L per L: the 63% almond share at the almond-milk figure, the rest at oat milk's (a Claude simplification) · NMPF (Circana data), Mekonnen & Hoekstra 2011

US bottled water 2025 88.4 billion liters 0.1–0.3× AI

16.8 billion gallons, preliminary (IBWA) · 1.39 L per L facility ratio — the one category below the AI range · IBWA 2026 progress report, IBWA 2024 benchmarking

The nine categories sum to 74.4 trillion liters a year — 97–238× the 2025 AI range, or 65–159× the projected 2026 range. The AI estimate covers the world; the drink totals cover the US only, so world drink totals would sit further right.

Two moving parts

The 2026 projection, and the shift from queries to tasks

Extrapolating to 2026

No one publishes a 2026 AI-water figure, so this is a Claude estimate. The Patterns model is mechanically AI power × a fixed water intensity, and AI power grew 2.45× from 2024 to 2025. Published rates for 2025→2026 run from ~1.3× (IEA accelerated servers) up to that 2.45× pace; scaling the 2025 range by a central 1.5× gives roughly 468.8 billion liters to 1.1 trillion liters for 2026. A UN University study puts the water footprint of data centers' 2025 electricity at 4.5 trillion liters with AI about 20% of that electricity — roughly 900 billion liters, 18% above de Vries-Gao's high estimate. At the 2.45× pace, the nine categories are still 40–97× the 2026 AI range.

A query is becoming a task

The published per-query figures describe one short text prompt (OpenAI's average, Google's median). Reasoning models emit far more tokens, and agents chain many calls per task, so the unit is shifting from query to task. At the default 2 mL scope, one beer is:

53,300 average ChatGPT queries
5,330 reasoning responses · 10× a query
3,550 agentic tasks · 15× a query

Multipliers from Epoch, Jegham et al., and Anthropic (multi-agent ≈ 15× a chat's tokens). Per-task water rises with the multiplier: at the default 2 mL scope one beer is 3,550 agentic tasks; at Mistral's 45 mL, 158.

Methodology

What's measured, and what isn't

The per-query figure is a range

Published water-per-query numbers span more than 100×, almost entirely because of where the boundary is drawn. OpenAI and Google report 0.32 mL and 0.26 mL. OpenAI's post states no boundary; Google defines its figure as on-site data-center cooling water only, so the page treats both as the cooling-only scope. Add the water used to generate the electricity and a modeled short GPT-4o query lands near 2 mL (Jegham et al., preprint). A life-cycle assessment of one 400-token Le Chat response — cooling, upstream electricity, and hardware manufacturing, no training — reaches 45 mL (Mistral). The selector shows all three and defaults to the middle scope — the closest match to the drinks' upstream accounting; the crop tables also count rain and pollution-dilution water, which have no AI analog. Even at 45 mL, one beer still equals more than 2,000 queries. Sam Altman's September 2026 comparison — 38,000 queries per California almond, cited from memory — pairs OpenAI's 0.32 mL figure with the full California almond footprint (Fulton et al. 2019: 12 L per kernel). On the middle scope the same almond is about 6,140 queries, or 9,660 on the global-average table this page uses. The "Almond" selector runs the numbers.

The AI annual figure

de Vries-Gao (Patterns, 2026) models AI worldwide at 312.5–764.6 billion liters consumed in 2025, built up from data-center disclosures rather than a directly reported total. LBNL separately puts all US data centers near 17 billion gallons on-site in 2023. A competing projection from Li & Ren reaches 4.2–6.6 billion m³ by 2027, but measures withdrawal, not consumption — a different quantity, so it is not charted here.

Drink footprints

Beer, wine, juice, coffee, tea, and milk use the canonical peer-reviewed crop and animal water tables (Mekonnen & Hoekstra 2011, 2012) — full green, blue, and grey water. Coffee and tea are set by bean and leaf mass, not cup volume. Almond and oat milk apply the same tables to what is in the carton — 2% almonds and 10% oats by weight, from the Almond Breeze and Oatly labels — so a richer recipe scales the figure in proportion. The widely cited Poore & Nemecek milk comparison (371 L almond, 48 L oat, 628 L dairy per liter) counts irrigation and processing water only, no rain, and does not publish recipes, so it is not charted with the crop-table figures. Soda uses the low end of the 169–309 L range (Ercin et al. 2011). Bottled water is a 1.39 L/L facility ratio and tap water is just the glass, so both understate next to the crop-based drinks. Spirits (spiritsEUROPE) and bottled water (IBWA) are the two industry figures; the ethanol crop-water tables (Gerbens-Leenes & Hoekstra 2009) imply 24–50 L for the spirit alone, above the 18 L used here.

US volumes

Each category uses its latest public total: beer and wine from NIAAA (2023), orange juice from USDA ERS (2022, the latest in the series), milk from USDA ERS (2025), coffee from USDA FAS (2024/25, by bean mass), bottled water and soda from IBWA (2025). Soda has no published total, so it comes from IBWA's own per-capita figure and implied population; IBWA's own market-share split (29% bottled water, 21% soda) independently gives about 12.2 billion gallons, within 1.5%. Two figures are soft: tea is priced by leaf mass from an industry import projection (Tea Association: 265–270 million lb for 2024), and plant-based milk is a retail-scan total (Circana, via NMPF) priced at the almond-milk footprint for its 63% almond share and the oat-milk footprint for the rest — a Claude simplification. Andy Masley's post prompted the question.

Sources

All linked

AI water

OpenAI / Sam Altman

Average ChatGPT query: 0.34 Wh and 0.000085 gallons (0.32 mL) of water; the post states no boundary. June 10, 2025.

Sam Altman on Sources, Sept 2, 2026

"For every 38,000 ChatGPT queries, that is the same amount of water that is used in the production of a single almond in California" — cited from memory. Quoted by Tom's Hardware.

Google Cloud

Median Gemini Apps text prompt: 0.24 Wh and 0.26 mL of water, defined as data-center cooling only. 2025.

Jegham et al. 2025

"How hungry is AI?" (arXiv preprint, v6 Nov 2025). Adds power-plant water: a short GPT-4o query (ChatGPT's default until Aug 2025), about 300 output tokens, lands near 2 mL; a 1,000-token response near 6 mL.

Mistral AI

Life-cycle assessment, ISO 14040/44-compliant and reviewed by the audit consultancies Resilio and Hubblo: 45 mL of water per 400-token Le Chat response, marginal inference only. July 2025.

Li, Yang, Islam & Ren — "Making AI less thirsty"

GPT-3 on 2023 hardware at US-average WUE: a 500 mL bottle per 10–50 medium-length responses; 2027 global AI withdrawal projected at 4.2–6.6 billion m³. Communications of the ACM 68(7), July 2025; arXiv copy linked.

de Vries-Gao, Patterns 2026

AI's modeled global water footprint reaches 312.5–764.6 billion liters in 2025.

LBNL 2024 data-center report

US data centers consumed 66 billion liters (17.4 billion gallons) on-site in 2023, plus ~800 billion liters via electricity at 4.52 L/kWh. The latest LBNL water figures; the June 2026 update covers electricity only.