Home/The Load
07

What does AI draw from the grid?

The Load

Data centres were a flat line on the grid for the best part of a decade, until the cloud and then AI. Few countries meter them. Ireland does: there they took 23% of the electricity in 2025.

A microchip wired by a thick cable to a transmission pylon of the same height

Ask about this section

  • How much of the world's electricity do data centres use today?
  • Which of these figures are measured and which are modelled?
  • Why do the per-query numbers from different sources disagree so much?

Twenty Years of Data Centres estimate

The share of a country's electricity that data centres take, since 2000: Ireland and the Netherlands metered, the rest modelled, one line per published study. Inside the view: How much, by region · What 2030 could look like.

Why did data-centre electricity use stay flat for so long before AI took off? →
485 TWh in 2025world data-centre electricity, modelled by the International Energy Agency (IEA)

Every series is one author's model or estimate, kept separate; only the Irish and Dutch series are metered. The 2030 figures are scenarios: what given inputs would produce.

Open Twenty Years of Data Centres →

Where the Load Sits measured

Data-centre electricity by region in the latest year, and the share of national demand it takes, with Ireland as the outlier.

Why is Ireland's data-centre share so much higher than anyone else's? →
23% in 2025of Ireland's metered electricity went to data centres (Central Statistics Office)

Only Ireland's and the Netherlands' figures are metered. Every other region is a modelled estimate by its author, and definitions of a data centre differ between them.

Open Where the Load Sits →

The Cost of a Token self-reported

Joules per generated token, measured on the graphics chip (GPU) for dozens of open models, and how much the answer grows once the whole building is counted. Inside the view: Energy per generated token, by model and GPU.

How much energy does one AI question use? →
0.24 Whper text prompt, Google's fleet median, by its own account

Benchmark figures are GPU-only under lab conditions. Facility energy depends entirely on the server-overhead and PUE (cooling and building overhead) assumptions shown as sliders. Google's median covers text prompts only. Method: Energy per token.

Open The Cost of a Token →

Tokens as Grid Load scenario

How many tokens (fragments of text) AI companies say they process, each in its own unit, and what those volumes would draw from the grid at three energy costs per token, next to Ireland for scale. Inside the view: Implied electricity per disclosure, TWh a year.

Why can't the Google and OpenAI numbers be added together? →
~8 TWh a yearfor the largest disclosure, median scenario, an estimate

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

Open Tokens as Grid Load →

Where the Electricity Comes From estimate

Three ways of counting the fuel behind data-centre electricity give three answers for the same place and year, each correct on its own terms: the grid mix, the IEA's marginal scenario and the companies' own claims.

Is AI running on gas or on renewables? →
22% in 2025gas share of the world's data-centre electricity, on the grid mix where they sit

Location-based shares are Ember's measured grid mix applied to each region's data-centre total (an estimate by the International Energy Agency, IEA; for Ireland, the figure metered by its Central Statistics Office, CSO). Marginal shares are an IEA scenario, market shares the companies' own claims.

Open Where the Electricity Comes From →
Next questionWhat gets built, and how fast? →The Build-out