What should I ask an AI vendor for to report the carbon footprint of my usage?
For whoever writes to the AI vendors on behalf of the sustainability team. Factor version soma-ai-ef-2026.11-dataset-v3.1.0. Factors last updated . Location-based grid averages, not renewable contracts. Estimates from published benchmarks.
Page updated
Part of the SOMA AI Emission Factors cluster.
Short answer
Ask for usage by model (tokens if possible), the region that served it, the facility's PUE, location-based and market-based electricity, water, and any emissions figure with its boundary. Each item moves your estimate up the data quality ladder or replaces an assumption with a fact. The email below can be pasted as it is. SOMA AI publishes the open factors you will apply to whatever comes back.
The checklist
- Usage by model. Input tokens, cached input tokens and output tokens per model for the period, or requests and active users if tokens are not available.
- Serving region. The country or cloud region that processed the requests, and whether it is fixed by contract or chosen by the vendor at run time.
- Facility efficiency. The power usage effectiveness of the facilities, and the hardware generation serving the model.
- Electricity. Location-based and market-based electricity figures for the service, with the contractual instruments and their boundary.
- Water. On-site cooling water and any off-site water behind the electricity, with the basis used.
- Vendor footprint statement. Any emissions figure for your usage, with its boundary, the method and any assurance opinion.
- Format and cadence. A dated export you can keep as evidence, and how often it will arrive.
Keep the vendor's reply, with its date, as inventory evidence. A refusal is also evidence. Record it, then estimate and say that you did.
Request email
Subject - Request for AI usage data for our carbon inventory Hello, We report greenhouse gas emissions for our purchased services and want to include our use of your AI service for the period [start date] to [end date]. Could you send us the following for our account. - Input, cached input and output tokens by model and month, or requests and active users if tokens are not available. - The country or cloud region that served these requests. - The PUE and hardware generation of the facilities involved. - Location-based and market-based electricity figures for the service, and the instruments behind the market-based one. - Water used for cooling and for the electricity, and the basis. - Any emissions figure you hold for our usage, with its boundary, method and assurance. If you cannot provide an item, please tell us so, and we will record that in our inventory as a data limitation and estimate it. Thank you, [Name], [Company]
Worked example, what each reply lets you claim
| The vendor sends | You can use | Activity | Factor |
|---|---|---|---|
| Token export by model, no region | Tier 2a, region declared as an assumption | 2 | 3 |
| Token export and serving region | Tier 2a, region confirmed | 2 | 3 |
| Active users and messages only | Tier 2b | 4 | 3 |
| An assured footprint figure for your usage | Tier 3, supplier-specific | n/a | 1–2 |
| Nothing | Tier 1, spend-based | n/a | 5 |
The region and the PUE matter because they change the factor, and the water answer feeds the water footprint. If a vendor gives no usage data at all, run the SOMA AI calculator on seats and spend and send the request again next quarter.
Common mistakes
- Asking for a carbon figure only. Vendors usually have usage data they can export, so ask for that first.
- Accepting a market-based number with no instruments or boundary, then reporting it as the only figure.
- Not recording the request, so the next reviewer cannot tell a missing answer from a missing question.
