What methodology should a company use to estimate the energy use of generative AI inference?
For sustainability leads choosing a method before they collect any data. 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
Use the best method your data allows, following the GHG Protocol's ranking of Scope 3 methods by how specific the data are. Supplier-specific data comes first, then token-based activity data multiplied by a published factor, then message or seat counts, and spend-based screening only where no usage data exists. Today no AI vendor we found gives a supplier-specific figure, so token-based factors are the practical standard where you can get billed tokens. SOMA AI publishes the open SAEF factors for that route.
The methods, in order of preference
| Method | Data needed | Activity score | Factor score | Acceptable when |
|---|---|---|---|---|
| Supplier-specific | The vendor's assured figure for your usage | n/a | 1–2 | Always, once a vendor offers one with a stated boundary |
| Token-based (Tier 2a) | Billed tokens by model and region | 2 | 3 | Whenever the API usage export exists |
| Seat or message-based (Tier 2b) | Active users, messages | 4 | 3 | Seat-licensed tools with no token export |
| Spend-based (Tier 1) | Invoice amounts | n/a | 5 | Screening, or small lines you will refine later |
The order follows the GHG Protocol Scope 3 Technical Guidance, which ranks its four methods for purchased goods and services, supplier-specific, hybrid, average-data and spend-based, by how specific they are to the supplier. It does not require the most specific method first. The hybrid case, where a vendor gives energy rather than a CO2e figure, sits beside the token-based row. Purchased AI inference is reported in Category 1, purchased goods and services. How SOMA's tiers map to the draft data types is set out in the data-tiers guide, and the equations are on the methodology page.
Location-based and market-based
The SAEF factors are location-based. They use the average carbon intensity of the grid where the model is served and take no credit for renewable certificates or power purchase agreements. A vendor's market-based claim can be reported beside the location-based figure if the vendor documents the instruments and the boundary. Do not subtract it from a location-based factor, because that mixes two accounting bases. Say which basis each line uses.
Worked example, three methods side by side
A company spends 200 EUR on an API (an illustrative price of 2 EUR per million tokens) for 100 million tokens, on a mid-size model served from US East - N. Virginia. A separate 50-seat chat tool has no token export.
| Method | Kg CO2e | Uncertainty |
|---|---|---|
| Token-based, API | 2.61 | class range up to 39% |
| Spend-based, same API | 19.2 | ±300% |
| Seat-based, 50 seats | 1 | dataset band on tokens |
The spend-based line is higher because money does not track electricity. It is a screening figure on a different boundary, never an upper bound, and not comparable with the token-based one. In SOMA's method paper, accepted for publication in Sustainable Production and Consumption, spend-based figures for the same purchase run 1.7 to 12.3 times a token-based estimate on deflated spend, as computed here, and 2.0 to 15.1 times undeflated, depending on price tier and region. Only the token-based figure tells you what to change. See how to cut the footprint, and why published estimates differ. The SOMA AI calculator takes any of the three inputs.
Common mistakes
- Starting with spend because the invoice is easy to find, then never replacing it.
- Comparing a spend-based figure with a token-based one as if they covered the same thing.
- Netting renewable certificates against a location-based factor.