Every number on this page has a source you can open. Where a number is estimated rather than measured, it says so. Where we have been wrong, that is here too, with the correction.
The counter on JSTJO shows 1 drop per answer. You should know exactly what that means before you trust anything else here.
Right now a drop is a unit, not a measurement. Every answer costs one drop whether it is a one-word reply or a long answer that ran ten web searches. Those two genuinely cost different amounts of water — by orders of magnitude — and the meter does not currently distinguish them.
So the drop counter is a fair-use allowance wearing a meter's clothes. Calling it a measurement would be a lie, and this project does not get to make that one.
Making drops vary with the actual tokens and searches behind each answer is the next thing being built. When it ships, this page changes and the change is logged at the bottom. Until then, the number below is what a typical answer costs, not what your answer cost.
Nobody — including OpenAI, Google, or us — can tell you the water a specific query pulled from a specific datacenter. That data is not published per request. Anyone quoting you a precise per-query figure is modelling it. So are we. The difference we can offer is showing the model.
Published per-prompt estimates differ by a factor of about two thousand. That is not because someone is lying. It is because they are counting different things.
| Estimate | Per prompt | What it counts |
|---|---|---|
| Google, median Gemini text prompt technical report, Aug 2025 |
0.26 mL | Onsite cooling only. Active chips, idle machines and cooling overhead — but not the water used generating the electricity. |
| OpenAI, Sam Altman “The Gentle Singularity”, June 2025 |
~0.32 mL | Methodology not published. “Average query” undefined. Training excluded. |
| UC Riverside, Li et al. Communications of the ACM, 2025 |
~2.2 mL | Onsite, US average across datacenters. |
| Washington Post × UC Riverside Sept 2024, 100-word email, GPT-4 |
~519 mL | Onsite plus the water used generating the electricity. Ranged 235 mL in Texas to 1,408 mL in Washington for the identical email. |
The dominant reason is the accounting boundary. The US Department of Energy's 2024 datacenter report found American datacenters consumed roughly 66 billion litres directly, and close to 800 billion litres indirectly through the electricity they used.
So a figure that counts only onsite cooling is leaving out roughly twelve times more water than it includes. That is not a rounding difference. It is the whole argument.
We use a mid-range coefficient and state the boundary alongside it. A typical JSTJO answer, counting onsite cooling only:
If someone tells you it is 0.26 mL, they are counting onsite only. If someone tells you it is half a litre, they are counting a long generation with the full electricity supply chain. Both can be honest. Neither is the whole picture on its own.
This section exists because a methodology page without one is marketing.
| Unknown | Why |
|---|---|
| Which datacenter served you | Water intensity varies enormously by region — the same email ranged 235 mL to 1,408 mL depending on location and grid mix. We do not know which one answered you. |
| Real-time cooling load | Evaporative cooling uses far more water on a hot day. Nobody publishes this per request. |
| Anthropic's per-query figure | JSTJO runs on Claude. Anthropic has not published a per-query water number, so our coefficient comes from independent research on comparable models, not from our own provider. |
| Training, amortised | Training GPT-3 alone consumed an estimated 700,000 litres onsite. Spreading that across queries requires knowing total query volume, which is not public. We exclude it. That makes our number lower than the truth. |
One thing we can promise about the number: it is arithmetic in the app, not something the model generated. A coefficient multiplied by a token count. It can be wrong — and it has been — but it cannot be hallucinated, and wrong in a consistent direction is the only kind of wrong that can be found and fixed.
True, and the comparison is usually made wrong — including by people arguing our side.
Beef's headline footprint is about 15,400 litres per kg. But 94% of that is green water — rain falling on pasture that was going to fall anyway. The blue water, actually withdrawn from a river or aquifer, is about 550 litres per kg.
| 1 kg of beef | Litres | Share |
|---|---|---|
| Green — rainfall on pasture and feed | 14,414 | 94% |
| Blue — irrigation, drinking, service water | 550 | 4% |
| Grey — dilution of pollutants | 451 | 3% |
Datacenter cooling is all blue water. So the only fair comparison is blue to blue, with the same boundary on both sides. Do that and a kilo of beef comes out somewhere between 1,000 and 2 million AI queries, depending entirely on whether you count electricity-generation water on both sides.
The low end of that range is not flattering to us. If you count the full supply chain on both sides, one kilo of beef is around a thousand heavy queries — which a regular user clears in a month. AI water use stops being trivially small the moment you compare like for like.
We are publishing that because it is true, not because it helps.
Every error found in a public number, what it was, and what changed. This list only grows. Nothing gets deleted from it.
Primary sources only. Where a figure is contested, both sides are listed.
| Figure | Source |
|---|---|
| Beef water footprint, green/blue/grey split | Mekonnen & Hoekstra, A Global Assessment of the Water Footprint of Farm Animal Products, Ecosystems 15:401–415, 2012 — PDF |
| 0.26 mL per Gemini text prompt | Google, Measuring the environmental impact of AI inference, Google Cloud Blog, August 2025 |
| ~0.32 mL per ChatGPT query | Sam Altman, The Gentle Singularity, June 2025. Not peer reviewed; methodology undisclosed. |
| ~2.2 mL onsite; up to ~50 mL with offsite | Li, Yang, Islam & Ren, Making AI Less “Thirsty”, Communications of the ACM 68:54–63, 2025 |
| ~519 mL per 100-word email | Washington Post with UC Riverside, September 2024 |
| US datacenter direct vs indirect water | Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report, LBNL-2001637, December 2024 |
| Google company water use, 2023–2024 | Google Environmental Report — company-disclosed, not modelled |
Spot-check us. Every figure above is either linked or named precisely enough to find in one search. If a number here does not match its source, tell us and it goes in the corrections log with your name on it.