How much energy does AI use?
AI uses a lot of electricity and water in aggregate, mainly in data centres for training and running models, and the demand is growing fast. A single question uses little, but billions of questions add up. Exact figures vary by model and are often not disclosed by the companies.
The full answer
AI's energy use is real and rising, and the numbers are uncertain.
Where it goes: training a large model takes a great deal of computing in data centres, and then inference, the running of the model for users, happens every time someone asks a question. At global scale inference is a large and growing share. Data centres also use water for cooling.
Per question vs in total: one chat request uses a small amount of energy, which can make it feel trivial. But billions of requests a day, plus image and video generation, which cost more, add up to a significant load on power grids.
Why there is no single number: energy per request depends on the model size, the hardware, the length of the answer and where the data centre gets its power. Companies rarely publish full figures, so estimates differ widely, and we do not quote one here.
What changes it: smaller and more efficient models, better chips, cleaner electricity and using AI only when it is needed. Using a smaller model for simple tasks is one practical step.