Chapter 09 · Sustainability & AISustainability & AI

AI carbon footprint

Meaning statusEstablishedSource recordDirect document linkedWhy these are different

Definition

The greenhouse-gas emissions attributable to an AI system across its life cycle: hardware manufacture, model training, and ongoing operation (inference).

References

ACL 2019 / arXivStrubell et al. — Energy and Policy Considerations for Deep Learning in NLP

This source provides part of the technical or institutional basis for the definition.

arXiv (Google)Patterson et al. — Carbon Emissions and Large Neural Network Training

This source supports the explanation of how the term is applied, measured or governed in practice.

FAccT 2022 / arXivDodge et al. — Measuring the Carbon Intensity of AI in Cloud Instances

This source supports the wider sustainability significance and context described in the entry.

arXivLuccioni et al. — Estimating the Carbon Footprint of BLOOM

This source supports the wider sustainability significance and context described in the entry.

Overview

What it means

Accounting follows established boundaries — embodied emissions of chips and servers, electricity consumed in training and serving, weighted by grid carbon intensity. Peer-reviewed studies (Strubell 2019; Patterson 2021; Dodge 2022; Luccioni 2022) established the method and showed results vary by orders of magnitude with scale, hardware and location.

How it is used

Organisations increasingly include AI use in scope 3 and ICT inventories; researchers publish per-model footprint estimates; and the EU AI Act obliges general-purpose model providers to document energy consumption. Estimates remain sensitive to allocation choices and data gaps.

Why it matters

AI's climate ledger has two columns — emissions caused and emissions potentially avoided. Without credible footprint accounting, neither the cost side nor the benefit side of that ledger can be managed, compared or regulated.