07What does AI draw from the grid?

Tokens as Grid Load: what disclosed AI use would draw

AI companies publish how many tokens they process (a token is a fragment of text, about three-quarters of a word): how much electricity would that take?

Google's latest disclosure (30 September 2025) works out to about 43 trillion tokens a day, counted as tokens across all its products (surface tokens), the largest published figure of any company.

Its separate API (developer) count (29 October 2025) is a different slice, about 10 trillion tokens a day.

OpenAI's latest API (developer) disclosure (6 October 2025) puts it at about 8.6 trillion tokens a day.

The series cannot be added: each company counts something different, so the chart keeps one line per company and definition and shows no total.

Take Google's disclosed surface tokens and multiply by an assumed energy per token: the result is roughly 0.4 to 91 TWh a year, 7.5 TWh at the median, an average draw of about 860 MW.

For scale, the IEA's modelled figure for all the world's data centres in 2025 is 485 TWh; Ireland's metered data-centre draw is the other reference line.

An illustrative estimate: a disclosed volume multiplied by an assumed energy per token. It measures no company's electricity use; read the whole range.

Disclosed volume, tokens per day within each definition
Log scale. Hover a point for the company's own words and the source.

Dotted lines are company figures, self-reported. A hollow point was converted from prompts or words with a stated factor.

Loading data…
Implied electricity per disclosure, TWh a year

Loading data…

How these numbers are made: Implied data-centre load.

Get an email when this section changes.