The science

AI carbon numbers you can defend.

Our method is peer-reviewed and published by Elsevier. The factors behind it are open data. Every number in AI4AI traces back to a published source that anyone can check.
Read the paper

Spend-based estimates miss by up to 12×.

Most inventories put AI inside a generic IT spend line. It can't tell one model or one region from another, and nobody can reduce what they can't see.

Spend-based, today

Euros × an IT-sector factor

  • 1.7 to 12.3 times the physical estimate
  • Same number for a small model and a frontier one
  • Same number in Sweden and in Singapore

Token-based, AI4AI

Tokens × a factor per model and region

  • Built from public benchmarks and grid data
  • Separates models, regions and providers
  • Every input sourced, versioned and open

How the method works

From a token count to a line in your inventory.

  1. 1

    Start from what you were billed.

    Token counts from the provider invoice. When a tool only shows seats or messages, the method estimates tokens from them and says so.

  2. 2

    Build the factor from public evidence.

    Benchmarked GPU energy per token, server overhead, data-centre efficiency and the grid of the cloud region. Then hardware, construction and model training on top.

  3. 3

    Report one line an auditor can follow.

    kg CO2e in Scope 3 Category 1, with electricity and water beside it. Every input is tagged as measured, published, assumed or derived.

Calculation chain from the paper: billed tokens become effective tokens, multiplied by a factor built from GPU energy, host, idle, PUE and grid intensity, plus embodied hardware, construction and training, giving kg CO2e in Scope 3 Category 1
Figure 3 of the paper. The calculation chain, coloured by the evidence behind each input.

The model and the region decide the number.

12×

between model classes

The same question costs about twelve times more carbon on a frontier model than on a small one.

9×

between cloud regions

The grid behind the data centre changes the figure as much as the model does.

0.039 kg

CO2e per million tokens

A mid-size model on the world-average grid. The starting point, not the answer.

Worked examples from the paper

8.5 kg

CO2e a year for a 200-person company

2.1 t

CO2e a year for 50 developers with AI coding agents, mid-size model

8.3 t

The same team on a frontier model

Start with what you have. Climb when you can.

The tier follows the data each tool gives you, mapped to the GHG Protocol data-quality hierarchy that ESRS E1 asks for.

  1. Tier 1

    Spend

    Euros spent times an IT-sector factor. A screening figure only.

  2. Tier 2b

    Seats and messages

    Tokens estimated from what the admin console shows.

  3. Tier 2a

    Exact tokens

    Token counts from the billing export, times the factor for that model and region.

  4. Tier 3

    Provider certified

    A figure the provider metered for your account. Used as is.

Less precise → more precise

Check every number yourself.

No black boxes. The paper, the factors and the code are public. Our figures are estimates from published benchmarks, not meter readings, and we say so on every one.

Journal
Sustainable Production and Consumption
Publisher
Elsevier
Ranking
Q1, peer-reviewed
Published
14 November 2026, vol. 70, pp. 320–347
DOI
10.1016/j.spc.2026.11.003
  • 3

    Model classes, small to frontier

  • 18

    Cloud regions with their own grid

  • 4

    Data tiers, mapped to the GHG Protocol

  • CC BY 4.0

    Open factors anyone can reproduce

Want to see it on your own AI use?