Home Writing Projects Setup Protocol 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 both data-center cooling and the power plants behind it, the same supply-chain accounting the drink figures use. One beer ≈ 53,300 queries; by OpenAI's cooling-only figure, 331,000. Pick a drink, shift the scope, or count by agentic tasks, then see how whole US drink categories compare with everything AI drank in 2025 — and what it is on track to drink in 2026.

One drink in ChatGPT queries

One beer
53,300
average ChatGPT queries

Enough water for 30 queries a day for 5 years.

Across accounting 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.

The eight categories sum to 75.6 trillion liters a year — 99–242x the 2025 AI range, or 66–161x 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.

Where this is heading

2026, and the shift from queries to agents

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 independently implies about 0.9 trillion L of AI water in 2025, near the same band. Even on the most aggressive growth, the drink categories still outweigh AI by tens of times.

A query is becoming a task

The published per-query figures are median single prompts. Reasoning models emit far more tokens, and agents chain many calls per task, so the honest unit is shifting. 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, but so does what each task does — a beer still covers thousands of even the heaviest ones.

Methodology

What's measured, and what isn't

The per-query figure is a range

Published water-per-query numbers span more than 100x, almost entirely because of where the boundary is drawn. OpenAI and Google report about 0.26–0.32 mL for data-center cooling alone. Add the water used to generate the electricity and an efficient 2025 model lands near 2 mL (Jegham et al.). A full per-response life-cycle assessment of a large model reaches 45 mL (Mistral). The selector shows all three and defaults to the middle scope — the boundary that matches the drinks' supply-chain accounting. Even at 45 mL, one beer still equals more than 2,000 queries.

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. It covers the whole world. 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. 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 is the one industry figure, and the crop literature suggests it undercounts.

US volumes

Each category uses its latest public total: beer and wine from NIAAA (2023), orange juice from USDA ERS (2021, the latest in the series), milk from USDA ERS (2024), 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. Tea is the softest number — an industry serving count (Tea Association). Andy Masley's post prompted the question.

Sources

All linked