What One 100-Word AI Answer Really Costs in Water and Electricity
You type a question, press send, and a neat 100 word answer appears in seconds. It feels free. It is not. Somewhere, a data centre just spent real water and real electricity to write it. How much exactly is one of the most argued-over numbers in tech, and this article gives you both sides.
The famous estimate: a bottle of water per answer
The Washington Post, working with researchers at the University of California, Riverside, estimated in 2024 that a single 100 word answer from GPT-4 consumes about 519 millilitres of water, just over one full bottle, and about 0.14 kilowatt hours of electricity, enough to keep 14 LED bulbs glowing for an hour or charge a smartphone roughly nine times.
Why water at all? AI chips run hot, and most data centres cool them the way your body does, by evaporating water. Here is the part almost everyone misses: only about one fifth of that water cools the data centre itself. Roughly four fifths evaporates miles away, at the power plant generating the electricity. AI’s water problem is largely an electricity problem wearing a water costume.
The other side: the companies say it is far less
An honest article must tell you that the AI companies report dramatically lower numbers. Google measured its median Gemini text prompt at just 0.24 watt hours of electricity and about 0.26 millilitres of water, roughly five drops. OpenAI’s chief Sam Altman has said an average ChatGPT query uses around 0.34 watt hours. Those figures are hundreds of times below the famous bottle-of-water estimate.
Who is right? Both, partly. The scary estimate modelled a long answer from the biggest, most power hungry model of 2024 at an average American data centre. The company figures measure a typical short prompt on newer, more efficient systems, and they are the companies grading their own homework.
The true cost of question sits somewhere in between and depends on which model you use and how long the answer is. The International Energy Agency adds a hopeful note: energy used per AI task is falling at a pace it calls unprecedented.
Why small numbers still matter
Because they multiply by billions. Training one older model, GPT-3, evaporated an estimated 700,000 litres of fresh water before answering a single user.
Data centres globally already consume an estimated 560 billion litres of water a year, and a single large facility can drink up to 19 million litres a day, as much as a town of up to 50,000 people. That is why residents from Oregon in the US to Spain have pushed back when data centres compete with them for water, even though individual question costs, at worst, one bottle.
AI’s convenience has a physical bill, and the bill is paid in megawatts and litres by whichever town hosts the data centre.
The Four Pillars of AI Infrastructure: Chips, Electricity, the Grid and Water