Examples
Worked examples
- Is an instance
A model card declaring '450 MWh training energy, 80 tCO2e at the training-region grid intensity'.
- Is an instance
A research paper appendix using CodeCarbon to log training-run emissions.
Counter-examples
Looks similar, but isn't
- Not an instance
A statement of GPU-hours alone.
- Not an instance
A vague 'this work used renewable energy' assertion.
Editorial commentary
Training carbon footprint is the greenhouse-gas emissions attributable to the one-off (or periodically repeated) process of training a model: the energy consumed by the GPU/TPU cluster over the training run, adjusted for data-centre power-usage effectiveness (PUE), multiplied by the carbon intensity of the electricity grid supplying that data centre at the time. Strubell et al. (2019) were among the first to put concrete, widely cited numbers on this for NLP models, prompting Schwartz et al.’s ‘Green AI’ (2020) framing, which argued for reporting efficiency, not just accuracy, as a first-class research metric.
Reproducible reporting requires disclosing: hardware type and count (e.g. “2,048 H100 GPUs”), training wall-clock duration, measured or estimated energy draw, PUE, and the carbon intensity of the specific grid and time period — a training run on a low-carbon grid (hydro, nuclear-heavy) can differ from the same run on a coal-heavy grid by an order of magnitude in emissions for identical compute. Tools that estimate or log this include CodeCarbon and ML CO2 Impact; some providers now publish their own training carbon-transparency figures directly, though methodology and completeness vary and are not independently audited in most cases.
Training vs. inference: why this is only half the picture
Training is a large, front-loaded, one-time (or infrequent) cost. Inference carbon footprint is the recurring, per-request cost of actually running the trained model — and a growing body of research finds that for widely deployed models, cumulative inference emissions overtake training emissions well within the model’s operational lifetime, in some high-traffic cases within months of launch. Reporting training emissions alone therefore understates a model’s true lifetime footprint; see compute FLOPs estimate for the underlying compute-scale figure both footprint calculations build on, and carbon accounting (research) for the general methodology this AI-specific application draws on.
References
- Strubell, Ganesh, McCallum, ‘Energy and Policy Considerations for Deep Learning in NLP’ (ACL 2019)
- Schwartz et al., ‘Green AI’ (Communications of the ACM, 2020)
Also known as
model training CO2 · training emissions
Machine-readable encodings
Use in your systems
<role vocab="credit"
vocab-identifier="https://casrai.org/dictionary/"
vocab-term="Training carbon footprint"
vocab-term-identifier="https://casrai.org/dictionary/term/training-carbon-footprint" />{
"@context": "https://schema.org",
"@type": "DefinedTerm",
"@id": "https://casrai.org/dictionary/term/training-carbon-footprint",
"name": "Training carbon footprint",
"identifier": "https://casrai.org/dictionary/term/training-carbon-footprint",
"description": "The total greenhouse-gas emissions, expressed in kilograms or tonnes of CO2-equivalent, attributable to training a machine-learning model, estimated from energy consumption and the carbon intensity of the electricity supply.",
"inDefinedTermSet": "https://casrai.org/dictionary/domain/ai-ml-research-outputs#set",
"url": "https://casrai.org/dictionary/term/training-carbon-footprint",
"sameAs": [
"model training CO2",
"training 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": "2026-08-22T15:54:57",
"inLanguage": "en-GB",
"isAccessibleForFree": true
}
