Every answer costs water.
AI isn't free — it just bills the planet.
Every time you ask ChatGPT, Claude, or Gemini a question, data centres pull freshwater to keep servers cool. This tool estimates how much water your AI sessions consume — so you can see the hidden cost behind every answer.
Calculate my usage →How much water did your AI session use?
Adjust the inputs below to match your typical AI usage. The calculator combines per-query water estimates from peer-reviewed research with regional data centre cooling efficiency to produce a transparent estimate — updated live as you change each setting.
Your session's water footprint
Based on your inputs, here is an estimated freshwater cost for this AI session. These figures combine direct cooling water and electricity-linked water use at data centres in your selected region — presented in terms you can actually feel.
~518ml
estimated for your session
Drinking glasses (250ml)
2.1
glasses
Fraction of a shower (60L)
0.9
% of one shower
Daily water for 1 person in water-stressed region (2L)
0.3
person's daily supply
If every day this month
15.5
litres per month
Monthly projection
If you ran this same session every day for a month, you'd consume approximately 15.5 litres of freshwater — just from AI queries.
At global scale
Your session is one drop. But hundreds of millions of people use AI every day. When you multiply a single session's water cost across the entire user base, the numbers become impossible to ignore.
100M
daily AI users (est.)
0.52L
avg water per session
3×
YoY growth in AI usage
21 Olympic swimming pools — if 100M users each send 10 queries today (0.1 billion litres).
That's freshwater pulled from rivers and aquifers — not recycled greywater — to cool the servers that power your answers.
How we calculate this
Transparency matters. Every number on this page comes from published research or disclosed corporate environmental reports — never from undisclosed internal data. Here is exactly how we arrive at your estimate.
How we calculate
For each query, we start with a base water estimate (millilitres per query) drawn from peer-reviewed research — primarily Making AI Less “Thirsty” by Li et al. (2023), which quantified the freshwater footprint of large language model inference. We then add the water embedded in the electricity consumed by that query, using the regional Water Usage Effectiveness (WUE) factor for data centres.
WUE measures how many litres of water a facility consumes per kilowatt-hour of IT energy. Typical values vary by climate and cooling strategy: the United States averages roughly 1.8 L/kWh, Europe around 1.2 L/kWh, broader Asia-Pacific regions about 2.1 L/kWh, and India approximately 2.5 L/kWh — reflecting hotter, drier conditions where evaporative cooling is more common. This calculator applies the WUE for the region you select.
The formula: ml = (mlPerQuery × queries × typeMultiplier) + (kWh × queries × typeMultiplier × WUE × 1000)
Query type multipliers account for how much extra compute different tasks require. A simple Q&A uses the baseline (×1). Long-form essays and reports demand more tokens and GPU time (×2.5). Code generation sits between the two (×1.8). Image synthesis is the most intensive (×6), since generating pixels requires sustained high-power GPU workloads and proportionally more cooling.
We distinguish two kinds of water use. Direct water is consumed on-site for evaporative cooling — water that evaporates from cooling towers and is lost from the local watershed. Indirect water is used upstream to generate the electricity powering the servers, especially at thermoelectric power plants. Both count toward the total freshwater footprint of an AI query.
Training a large language model from scratch can consume millions of litres of water — orders of magnitude more than any single inference session. This tool measures inference only (the queries you run day to day), because that is what most users can control and because per-query inference estimates are better supported by published data. Training costs are a separate, much larger discussion.
Sources
Making AI Less "Thirsty"
Li et al., UC Riverside & UT Arlington (2023)
Primary source for per-query water estimates
Google 2023 Environmental Report
Google LLC (2023)
Data centre WUE and total water withdrawal figures
IEA — Data Centres and Data Transmission Networks
International Energy Agency (2023)
Global energy and water context
Disclaimer: These are estimates. No AI company publicly discloses per-query water data. Every figure links to its source. Treat results as directional, not precise measurements.
Frequently asked questions
Common questions about AI water usage, data centre cooling, and how this calculator estimates your footprint.
How much water does AI use per query?
Peer-reviewed research estimates that a single ChatGPT-style query can consume roughly 10–50 millilitres of freshwater when you combine direct data centre cooling with the water embedded in the electricity that powers the servers. The exact figure depends on the model (larger models use more compute), the type of task (a short answer vs. code or image generation), and where the data centre is located. This tool uses published per-query estimates from Li et al. (2023) as a baseline, then adjusts for your inputs.
Which AI model uses the most water?
Among the models in this calculator, GPT-4o-class systems generally carry the highest per-query water footprint because they require more compute and electricity per response. GPT-3.5-style models are lighter, while Claude and Gemini sit in between based on disclosed environmental data and published estimates. Image-generation workloads can dwarf text-only queries regardless of model, since synthesising pixels demands far more GPU time and cooling load than a brief text reply.
Why do data centres use water?
AI models run on dense clusters of GPUs and CPUs that generate enormous heat. Data centres must remove that heat continuously to prevent hardware failure. Many facilities use evaporative cooling — essentially allowing water to evaporate off cooling towers or similar systems — which is energy-efficient but consumes freshwater directly. Even centres that rely on air cooling still draw water indirectly through the power grid, because thermoelectric power plants use water for cooling and steam generation.
Is AI water usage a real environmental concern?
Yes. Global data centre water withdrawal is already measured in billions of litres per year, and AI workloads are growing faster than general internet traffic. Water stress is acute in regions such as the western United States, parts of India, and other areas where major cloud providers operate large campuses. While any single query is small, billions of daily queries aggregate into meaningful freshwater demand — often in places already facing drought or groundwater depletion.
How does data centre region affect AI water usage?
Region matters because of Water Usage Effectiveness (WUE) — the litres of water consumed per kilowatt-hour of IT energy at a facility. Hot, dry climates tend to rely more heavily on evaporative cooling and therefore have higher WUE values. A query routed through a US data centre (~1.8 L/kWh) will carry a different water footprint than the same query in Europe (~1.2) or India (~2.5). This calculator applies regional WUE multipliers so you can see how geography changes the hidden cost of the same prompt.
What is WUE (Water Usage Effectiveness)?
Water Usage Effectiveness (WUE) is a data centre sustainability metric defined as total facility water consumption divided by IT equipment energy use, expressed as litres per kilowatt-hour (L/kWh). A WUE of 1.8 means the site uses 1.8 litres of water for every kilowatt-hour of compute power. Lower WUE is better. Operators report WUE in environmental disclosures; researchers use it to compare cooling efficiency across regions and to estimate the water embedded in cloud workloads such as large language model inference.