Electricity use of a text question to an AI

Reference value · a typical question · independent estimate, February 2025

A text question to an AI assistant consumes about 0.3 Wh of electricity in the servers, according to Epoch AI's February 2025 estimate for a GPT-4o answer of about 500 tokens.

An order of magnitude for a short question: a long document or reasoning mode multiplies it.

See the source

A Linky electricity meter: a question is counted in watt-hours. Photo Benoît Prieur, Wikimedia Commons, CC0

Definition and scope

Electricity consumed by the servers to answer a typical text question put to a large assistant, about 500 tokens produced. This is the value that the hub and the blog use for all their calculations.

What is included

The GPUs that compute the answer
The rest of the server and the cooling, through a coefficient

What is excluded

Training the model
Server manufacturing
The user's device and the network

How the value is obtained

Epoch AI estimate for GPT-4o, based on the assumed size of the model, the number of tokens produced and the efficiency of the GPUs. The 2 only values published by providers bracket it: 0.24 Wh measured by Google, 0.34 Wh stated by OpenAI without a method.

Limits and uncertainties

An estimate, not a measurement: the model and its hardware are not published.
The real average depends on the length of the exchanges and the share of reasoning mode, which can multiply the energy by 2 to 30.