The CO2 cost per query exists. It is rarely disclosed.
Literal meaning: The total greenhouse gas emissions attributable to an AI system — calculated per query, per training run, or per model.
Origin: The concept sharpened analytically after Strubell et al.’s 2019 measurements, and was extended by researchers at Google and elsewhere who began publishing emissions figures for specific models. Figures vary enormously depending on the energy source of the data centre — a model trained on hydroelectric power looks very different from one trained on coal.
The CO2 cost of AI at every stage — from training to deployment to each individual query — measured but rarely published.
The Appeal: The term gives something previously invisible a number. Numbers enable comparison, accountability, and pressure. Some companies have responded to carbon footprint scrutiny by publishing emissions data and investing in renewable energy. The frame works.
The Friction: The frame also has a design flaw. Greenwashing — sustainability rhetoric deployed as marketing without changing underlying behaviour — exploits the same vocabulary. A company can publish a carbon footprint figure while simultaneously tripling its energy consumption. More structurally: Inference Cost — the energy consumed per individual query — is almost never included in published figures. The iceberg stays submerged. Accounting choices are political choices. Once you see the boundary of what is counted, you can ask who drew it.
Why This Matters: The term teaches a specific reading practice: ask not just what is the number but what does the number include. A disclosed footprint is not the same as a small one.
Related terms: AI Energy Consumption · Training Run · Inference Cost · Greenwashing · Externalized Costs · Digital Colonialism · Sacrifice Zones
Read more:
- Energy and Policy Considerations for Deep Learning in NLP — Strubell, E., Ganesh, A. & McCallum, A. (2019). ACL Chasing Carbon: The Elusive Environmental Footprint of Computing — Gupta, U. et al. (2021). IEEE HPCA