LLM Training Run Carbon Intensity Calculator

Combine GPU runtime, electrical efficiency, grid intensity, and mitigation actions to quantify the carbon footprint of a language-model training run.

Sum of accelerator count × runtime for the entire training job.
Sustained electrical draw per accelerator at the utilisation achieved.
Location-based emission intensity of delivered electricity.
Power usage effectiveness for the data center during the run.
Defaults to 0%. Represents bundled renewables or credits that zero out matching load.
Enter verified offsets retired for this run; defaults to 0.
Used to report emissions per million tokens; leave blank if unknown.

Indicative estimator; complement with lifecycle assessments and contract-specific emission factors for formal ESG reporting.

Examples

  • 512 GPU-hours, 0.55 kW per GPU, 0.45 kg/kWh grid, 1.28 PUE, 35% renewables, no offsets, 0.24B tokens ⇒ 105.43 kg net, 0.206 kg/GPU-hr, 0.439 kg/Mtoken
  • 2,048 GPU-hours, 0.65 kW, 0.50 kg/kWh grid, 1.35 PUE, 80% renewables, 1,500 kg offsets ⇒ 0 kg net emissions

FAQ

Which emission factor should I use—location-based or market-based?

Use the factor aligned with your reporting boundary. Location-based factors reflect the regional grid mix, while market-based factors incorporate contractual instruments such as RECs. Enter the one you report publicly and set the renewable percentage to the share of load met with bundled clean supply.

Does the calculator handle dynamic PUE during the run?

It assumes a constant PUE. If your facility publishes hourly or seasonal PUE, average the values weighted by GPU-hours in each interval before entering the figure.

How are offsets applied?

Offsets are subtracted directly from gross emissions in kilograms of CO₂e. Enter only retired credits with verifiable serial numbers so the net total reflects actual mitigation.

Can I translate the result into emissions per training token?

Yes. Provide the total tokens processed in billions to display emissions per million tokens, which is useful for comparing runs or pricing carbon-aware API access.

Additional Information

  • IT energy equals GPU-hours multiplied by sustained power per accelerator.
  • Total facility energy scales IT energy by PUE to account for cooling, power conversion, and overhead systems.
  • Net emissions subtract offsets from gross emissions after adjusting for renewable supply matched to the run.

Related calculators