How do I estimate the emissions of our AI model training and LLM API usage?

Question

Senseible · Accepted Answer

For GPU training, multiply GPU-hours by each board's power draw, then by the data centre's PUE, then by the grid factor of the region. For third-party LLM APIs, most providers publish no per-token factor, so use the provider's figure if one exists, otherwise a spend-based estimate labelled as such. GPUs in your own office or server room, on your electricity bill: Scope 2. Rented cloud GPUs and model APIs: Scope 3 Category 1, purchased services. GPU servers you buy: their manufacturing emissions sit in Scope 3 Category 2, capital goods. Take GPU-hours from the job logs or cloud bill. Multiply by power per GPU in kW. The rated board power is an upper bound; measured draw from monitoring tools is better. Add the rest of the server (CPU, memory, networking) if you can measure it, or state that you left it out. Multiply by the facility PUE, from the provider or a stated assumption. Multiply by the grid factor for the hosting region: 0.710 kg CO2 per kWh for India-hosted compute (CEA, version 21.0). Most model providers do not publish emissions per token or per request, and the figures that do appear use different boundaries. Ask your provider for your account's figure first. If none exists, take the pre-tax spend from the invoices and apply a spend-based factor for computer or data processing services, converted to the factor's currency and year. Label it spend-based with wide uncertainty, and replace it when the provider gives data. GPU-hours, power assumption, PUE and grid factor, each with its source. Whether the figure is location-based or market-based. Which API costs are spend-based estimates.

How do I estimate the emissions of our AI model training and LLM API usage?

Which scope it falls in

Training runs: the energy method

API usage when the provider publishes nothing

What to disclose

Worked example: a fine-tuning run in an Indian region

What Senseible does here