What “training” means
An AI model like GPT-3 is not born knowing anything. Before it can chat with anyone, it must be trained: computers read enormous amounts of text from books and the internet, over and over, for weeks or months without stopping, adjusting billions of internal settings until the model learns language.
Think of it like a student cramming day and night for months before their very first exam. All that cramming happens on thousands of hot, power-hungry chips running non-stop.
Where the water comes in
Those chips generate tremendous heat, and the data centre cools them mostly by evaporating water, the same way your body sweats. Evaporated water rises into the air and is gone from the local supply, it does not go back down the drain to be reused.
Researchers at the University of California, Riverside calculated that just the training phase of GPT-3, done in Microsoft’s American data centres, evaporated an estimated 700,000 litres of clean, fresh water. That is the cost before the model had answered even one question from one user, like a restaurant burning through a full tank of cooking gas before serving its first customer.
How much is 700,000 litres, practically?
Some everyday comparisons make it real. It is about 1.4 million half-litre water bottles. Using India’s standard of about 135 litres per person per day for all needs, bathing, cooking, drinking, washing, it is what one person would use in roughly 14 years, or a family of four in about three and a half years.
The researchers themselves compared it to the water needed to manufacture around 370 BMW cars.
Note
This was a 2023 estimate for an older model, and it counted only training. Newer data centres increasingly use closed loop cooling that recycles water, so newer models may differ. But an AI’s water bill starts running long before you type your first question.