Home/Method notes/Implied data-centre load

Method note · scenario

Implied data-centre load

What it measures

The average electricity load an AI provider's tokens, each roughly three-quarters of a word, imply once an energy cost per token is applied.

Inputs

  • Published token-volume disclosures per AI provider, converted to tokens per day within each disclosure's own definition
  • The energy per token at each scenario in the assumption set, worked out in the Energy per token note

Method and weighting

Each provider's disclosed volume of tokens is converted to tokens produced in a day, within that disclosure's own definition of what is counted; a disclosure given in prompts or words is converted using a stated factor.

The energy per token used here is worked out in the Energy per token note, for each scenario in the assumption set stated on the view.

Tokens per day are scaled up to a year and multiplied by that energy per token to give the year's energy, and dividing the year's energy by the hours in a year gives the average load.

For scale, the view sets that load beside the modelled figure of the IEA (the International Energy Agency) for all the world's data centres.

Window

As of each disclosure's own date, stated beside it; the assumption set is curated and dated. Data last refreshed on .

Known limits

  • A disclosure converted from prompts or words to tokens carries the error of the stated conversion factor, and the view marks it as converted.
  • The load is an average, and a data centre's draw varies through the day.

Used on

Sources: Epoch AI, IEA. Each is described on the Data sources page.