How much water does AI use?
A typical AI chat prompt uses between 0.3 mL and 50 mL of water (five drops to a shot glass), depending on what you count. A year of daily ChatGPT use lands between a glass of water and a few bathtubs. US data centers consumed about 17 billion gallons directly in 2023, projected to roughly double by 2028.
That spread is not sloppy science. The low numbers count data-center cooling for one query; the high numbers add the water behind electricity generation, model training and hardware manufacturing. Every number on this site carries a scenario label: Low (cooling only), Mid (+ electricity), High (full lifecycle). You always know which question is being answered.
Estimate your own AI water use
frontier class · 2 Wh per 1,000 tokens (mid benchmark)
The same prompt can be “five drops” or “a bottle of water” depending on what you count: just data-center cooling (Low), the water behind the electricity (Mid), or training and hardware too (High).
Read the methodology16.9gallons
= 63.9 liters
ChatGPT (standard), 25 prompts/day · per year, Mid scenario
Energy
12.8 kWh
Carbon
4.5 kg CO₂e
Why "five drops" and "a bottle of water" are both right
The two figures that frame the public argument:
| Figure | Who | What it counts |
|---|---|---|
| ~0.32 mL per query | Sam Altman, OpenAI (2025) | on-site cooling (our Low) |
| 0.26 mL per median prompt | Google's Gemini disclosure (2025) | on-site cooling, median prompt (Low) |
| ~45 mL per 400-token response | Mistral's lifecycle analysis (2025) | full lifecycle (our High) |
| ~519 mL per 100-word email | Washington Post with UC Riverside (2024) | cooling + electricity, at a high 2024 energy estimate for GPT-4 (our Mid boundary) |
Between the operator figures and Mistral's lifecycle figure sits the Mid scenario: cooling plus the water consumed generating the electricity, which on the US grid runs around 4–5 liters per kWh. Mid is our default because the electricity is real and attributable, while lifecycle amortization involves more modeling choices.
Withdrawal vs. consumption: the quiet 4× swing
A data center (or a power plant) withdraws water when it takes it from a river or aquifer, and consumes it when the water doesn't come back, usually because it evaporated. A facility can withdraw a large volume, warm it slightly and return nearly all of it; another can withdraw little and evaporate most of what it takes. Headlines routinely cite withdrawal figures (bigger) where consumption is the ecologically relevant quantity for a stressed basin, or the reverse, when the point is reassurance. Every figure on this site is consumption unless labeled otherwise, matching the convention in the UC Riverside and LBNL work. When two published numbers for the same facility differ by 3–4×, this distinction is usually why.
The aggregate picture
Individual prompts are small. Fleets are not:
- US data centers used about 176 TWh of electricity in 2023 (~4.4% of national consumption) and 17 billion gallons of water directly (LBNL, 2024). The 2028 projections roughly double both.
- Counting the water behind their electricity multiplies the footprint several-fold (Siddik et al. 2021).
- Google alone consumed about 7.7 billion gallons across its data centers in 2024, up from the 6.1 billion still widely quoted (that was 2023).
Two things are simultaneously true: your personal chatbot habit is hydrologically tiny, and a hyperscale campus in a drought-stressed county is a real local water user. The first fact is the calculator above; the second is our data center map.
How AI compares with the rest of your day
Using the High (full lifecycle) boundary throughout (the most generous accounting, so these are ceiling comparisons), with Water Footprint Network figures for the foods:
| One unit of… | Water, full lifecycle |
|---|---|
| AI chat prompt (High, Mistral LCA) | ~45 mL |
| Cup of coffee, including growing the beans | ~130 L |
| Hamburger (150 g beef patty), including feed | ~2,500 L |
| Pair of cotton jeans | ~7,500 L |
| One day of an average American's home water use (EPA) | ~310 L |
Fifty prompts a day for a year comes to roughly 800 liters at High (about six coffees' worth of embedded water) or 3–8 liters at Low. This is not an argument that the industry's totals don't matter; it is an argument that per-user arithmetic and infrastructure siting are different questions, and the second one is where the substance is.
Training vs. inference
Training is a one-time bulk cost: GPT-3's training was estimated to evaporate ~700,000 liters on site (about 5.4 million liters including the water behind its electricity); Mistral reported 281,000 m³ for Mistral Large 2's training plus its first 18 months of use. Spread over billions of subsequent queries, training typically adds less than the electricity term, which is why it only appears in our High scenario.
How to read an AI water headline
Four questions decode every AI-water claim you will ever encounter:
- What boundary? Cooling only, + electricity, or full lifecycle? This alone moves the answer ~100×. A number without a boundary is not yet a fact; it is a vibe with units.
- Withdrawal or consumption? If the article doesn't say, assume the larger-sounding one was chosen.
- Which model, what year? A GPT-3-era estimate applied to a 2026 efficient-tier model overstates by an order of magnitude; a median-prompt figure applied to a reasoning workload understates similarly.
- Median or mean, and of what? Operators report medians of short consumer prompts; researchers model specific, longer tasks. Both are honest; they are not the same number.
A worked example: "ChatGPT uses a bottle of water per conversation" decodes to UC Riverside, cooling + electricity (our Mid boundary), GPT-3-era hardware, 10–50 responses per 500 mL bottle: a defensible 2023 estimate for older, less efficient hardware, not a description of what evaporated while you asked for a recipe. Equally, "a few drops per prompt" decodes to operator-reported, Low (cooling only), median short prompt: also true, and also not the whole picture.
What would change these numbers
Disclosure, mostly. Google publishing its per-prompt figure moved this site's Gemini coefficient the week it happened. Efficiency gains (better chips, better cooling, cleaner grids) push per-query numbers down every year, while volume pushes aggregate numbers up. Both can be true at once; the methodology page tracks every coefficient and source we use.