Examples
Worked examples
- Is an instance
A model card reporting '~0.5 gCO2e per 1k-token request at the EU-West-1 inference region'.
- Is an instance
An LLM-serving provider's quarterly carbon-impact report.
Counter-examples
Looks similar, but isn't
- Not an instance
Training-only emissions estimate.
- Not an instance
An undated 'one Google search uses X joules' citation.
Editorial commentary
Inference carbon footprint is the greenhouse-gas emissions attributable to running a trained model to serve requests, rather than to training it. Estimating it requires per-request figures — token counts processed and generated, the model’s active parameter count (for a mixture-of-experts model, only a fraction of total parameters are active per token), hardware efficiency, and the carbon intensity of the grid supplying the inference data centre — multiplied by request volume.
This is where the current research finding worth stating plainly comes in: for a widely deployed model, the training run is a large but one-off cost, while inference recurs on every request served for the model’s operational lifetime. Patterson et al. (2021) and Luccioni & Strubell’s ‘Power Hungry Processing’ (FAccT 2024) both document this dynamic; secondary analyses aggregating provider-scale usage figures have put the point at which cumulative inference emissions for a very high-traffic, frontier-scale deployment match the original training run’s emissions at only a few months after launch — a figure that is deployment-specific, not a universal constant (a lightly used model may never cross that line), but the general dynamic it illustrates — inference dominating training over a widely used model’s lifetime — is corroborated across multiple independent studies, not a one-off finding.
What complicates measurement
Inference-time footprint is much harder to report reliably than training footprint: request volume and mix (short vs. long generations, cached vs. fresh computation) change continuously after launch, most providers do not publish per-query energy figures, and third-party estimates therefore rely on published hardware specifications and modelled utilisation rather than direct measurement. Reporting practice in this area is genuinely still maturing, not just under-disclosed by convention.
See also
See training carbon footprint for the front-loaded half of the same lifecycle, compute FLOPs estimate for the compute-scale figure both draw on, and carbon accounting (research) for the general accounting methodology.
References
- Patterson et al., ‘Carbon Emissions and Large Neural Network Training’ (arXiv, 2021)
- Luccioni, Strubell, ‘Power Hungry Processing’ (FAccT 2024)
Also known as
model inference CO2 · serving emissions
Machine-readable encodings
Use in your systems
<role vocab="credit"
vocab-identifier="https://casrai.org/dictionary/"
vocab-term="Inference carbon footprint"
vocab-term-identifier="https://casrai.org/dictionary/term/inference-carbon-footprint" />{
"@context": "https://schema.org",
"@type": "DefinedTerm",
"@id": "https://casrai.org/dictionary/term/inference-carbon-footprint",
"name": "Inference carbon footprint",
"identifier": "https://casrai.org/dictionary/term/inference-carbon-footprint",
"description": "The greenhouse-gas emissions associated with serving inference requests from a deployed model, typically expressed per-request (e.g., gCO2e per query) or in aggregate (kgCO2e per month).",
"inDefinedTermSet": "https://casrai.org/dictionary/domain/ai-ml-research-outputs#set",
"url": "https://casrai.org/dictionary/term/inference-carbon-footprint",
"alternateName": [
"model inference CO2",
"serving emissions"
],
"license": "https://creativecommons.org/licenses/by/4.0/",
"publisher": {
"@id": "https://casrai.org/#organization"
},
"author": {
"@id": "https://casrai.org/#editorial-team"
},
"datePublished": "2026-05-21T02:22:51",
"dateModified": "2026-09-05T17:31:45",
"inLanguage": "en-GB",
"isAccessibleForFree": true
}
