How Much Water Does Your Chatbot Actually Use? The Numbers Behind the Headlines

M By Marcus Webb · July 10, 2026

Somewhere in your feed this month, you have probably seen both of these claims: that every ChatGPT question guzzles a bottle of water, and that AI’s environmental impact is a rounding error not worth thinking about. They cannot both be right.

Having spent a fair amount of time in the sources behind these claims, I can report the unsatisfying truth: the honest answer sits in the middle, and the reason the debate feels so broken is that the two sides are usually not measuring the same thing.

This article walks through what the numbers actually say, where they come from, and what they mean for a professional who uses AI tools every day. No affiliate links in this one. Nothing to sell. Consider it background reading for the tools I do review.

The claim under examination

The most viral version of the water claim traces back to a 2023 research estimate that a ChatGPT conversation of 20 to 50 questions consumed roughly 500ml of water. Repeated through a few rounds of social media compression, this became “every prompt is a bottle of water.” Meanwhile, sceptics point out that a single hamburger supposedly uses 1,700 litres, making AI’s usage trivial by comparison.

Here is the problem: both figures are technically defensible and both are misleading without context. The burger comparison and the bottle comparison are counting different kinds of water.

The distinction that changes everything: blue water

Water footprint research splits water into three categories, and this split is doing almost all the work in the debate.

Green water is rainwater. It falls on farmland and would mostly evaporate whether or not anything was planted there. When you read that a beef burger takes 1,700 litres, around 90 per cent of that is rain landing on pasture. That water was never in a pipe, never in a reservoir, and was never available for anything else.

Blue water is the water that matters for scarcity arguments: freshwater actually drawn from rivers, lakes, and aquifers. Strip the rainwater out of the burger and its blue water footprint drops to roughly 6 litres. A cup of coffee falls from about 140 litres total to roughly 4 litres of blue water.

Grey water is a notional volume representing pollution dilution, which most careful analyses set aside entirely.

Data centre water use is essentially all blue water. It comes from municipal supplies and evaporates in cooling towers, or it evaporates at the power plant generating the electricity. So when someone compares AI’s water use to a hamburger’s 1,700 litres, they are comparing AI’s blue water to the burger’s mostly-green water. It is not a fair fight in either direction.

What a prompt actually costs

The most rigorous public estimates come from EcoLogits, an open-source project that models the energy, carbon, and water behind AI models based on their parameter counts and output lengths. Independent researcher Andy Masley has built the most useful presentation of this data I have found, a calculator that puts your personal AI usage in context against your actual footprint. His methodology page is exemplary: every figure sourced, every assumption stated, uncertainty ranges included.

The rough shape of the numbers, using EcoLogits’ benchmark of a typical 400-token chatbot reply:

A standard chatbot exchange consumes water on the order of millilitres, not litres. Single-digit millilitres for smaller models, more for the largest reasoning models. The “bottle of water” framing was built on older estimates, longer conversations, and worst-case data centre locations.

Where it gets more interesting is at the heavy end. A full AI coding session, the kind where an agent writes and edits code across many steps, runs to around 100,000 output tokens on EcoLogits’ benchmark. That is 250 times a chatbot reply. Masley notes that building his own calculator page took roughly 250,000 tokens of output, about the energy of running a dishwasher, or playing a PS5 for five to six hours.

That reframing matters for professionals. Your twenty daily questions to a chatbot are not the story. Heavy agentic workloads are where individual usage starts to register, and even then, “a dishwasher load” is the right mental unit. Not a swimming pool.

The comparison that actually clarifies

Here is the context that finally made the numbers click for me. Thermoelectric power plants evaporate roughly 1.8 litres of freshwater per kilowatt-hour generated, according to the US National Renewable Energy Laboratory. The average American home draws about 29 kWh of electricity per day.

Run the arithmetic and an ordinary house consumes roughly 50 litres of blue water per day through its electricity use alone, before anyone opens a tap. The same power plant cooling that makes up much of AI’s water footprint is embedded in every appliance you own.

Your air conditioner does not get headlines. The data centre does. The physics is the same.

So is the criticism wrong?

No, and this is where I part ways with the full dismissal crowd.

Individual prompts are cheap. Aggregate demand is not. Data centre buildouts are causing genuine local water stress in specific regions, particularly where facilities were sited in water-scarce areas. Grid-level electricity demand from AI is real and growing, and the carbon cost of a prompt varies almost fourfold depending on whether the grid behind it runs on hydro or coal. The average grid intensity ranges from around 125g CO2e per kWh in the UK to 580g in China.

The honest framing is this: your personal AI usage is environmentally trivial, and the industry’s aggregate trajectory is a legitimate infrastructure question. Both things are true. The debate goes wrong when someone uses the first fact to wave away the second, or uses the second to guilt individuals about the first.

What this means in practice

If you use AI tools professionally and the environmental question has been nagging at you, three takeaways from the actual numbers:

Your daily chatbot usage sits somewhere below your morning coffee in blue water terms. Adjusting your prompt count is not a meaningful environmental lever. If you want to cut your footprint, the big levers remain what they have always been: how you travel, how you heat your home, what you drive.

Heavy agentic and coding workloads are a different tier, closer to running appliances. Worth knowing, not worth panicking about.

If you want to see your own numbers rather than take anyone’s word for it, Masley’s calculator lets you enter your actual daily usage and see it against your real-world footprint. It is the single best resource I have found on this topic, and he has released the whole thing into the public domain.