Token-based method proposed for counting AI inference in corporate carbon inventories
An arXiv paper argues purchased AI inference belongs in Scope 3 Category 1 and offers a token-based accounting method with an open dataset, while noting accuracy against metered energy is untested.
In plain words
Companies that pay to use AI models currently struggle to count the emissions those services cause. The paper proposes calculating a footprint per million billed tokens, using a factor that reflects serving electricity, hardware, data-centre construction and disclosed training runs. The authors stress the method is traceable but has not been checked against metered energy.
What was announced
The paper, posted on arXiv, addresses a gap in corporate greenhouse gas accounting: purchased AI inference is usually omitted from Scope 3 Category 1 inventories or estimated with spend-based factors, which the author says give 1.7 to 12.3 times the physical estimate and cannot distinguish between models or regions. The proposed method defines a cradle-to-gate boundary, four tiers of activity data, an effective-token basis, and a four-term factor covering serving electricity, embodied hardware, data-centre construction and disclosed final training runs, plus a rule that resolves any model identifier and region to a factor.
The reported footprints vary widely: twelve times between model classes and nine between regions. The mid-class factor is 0.039 kg CO2e per million billed tokens at a world-average grid, described as the low end of a structural range, since a less efficient serving configuration raises it 2.8 times and broader training accounting gives 1.5 to 2.9 times it. Worked examples include 8.5 kg CO2e a year for a 200-person firm and 2.1 tonnes for fifty developers using agentic coding tools on a mid-class model, or 8.3 tonnes on a frontier-class one. The paper states accuracy against metered energy is untested.
What it does not show
- The evidence is the paper's abstract only; the full methodology, data and peer review are not covered.
- The author states accuracy against metered energy is untested, so the factors are not validated by measurement.
- Figures such as the worked examples come from the paper itself and have no independent confirmation.
For security and IT teams
- Scope
- The method targets purchased AI inference within Scope 3 Category 1 of a corporate greenhouse gas inventory, with a rule mapping any model identifier and region to a factor.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.