SOMA

    How do I calculate the carbon footprint of my company's AI usage?

    For sustainability and finance teams building a first AI line for their carbon inventory. 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

    List the AI tools, collect usage for a period, map each model to a size class and a serving region, multiply by a factor per million tokens, add the lines up and report them as Scope 3 estimates with their assumptions. With the open SAEF factors, a million tokens through a mid-size model is 0.00714 kg CO2e from EU North - Sweden and 0.0261 kg from US East - N. Virginia. SOMA AI publishes these factors, and the free calculator applies them to your tools.

    Steps

    1. Scope the tools. List every AI service the company pays for or lets staff use. Include API accounts, seat-licensed assistants such as ChatGPT Enterprise or Microsoft 365 Copilot, coding assistants, and AI features inside other software. Give each an owner and a reporting period.
    2. Collect usage data. For each service take the best activity data it offers. Billed tokens from the API usage export are best. Active users and message counts from an admin dashboard come next. Spend on the invoice is the last resort.
    3. Map each model to a size class and a region. Assign each model to a small, mid-size or frontier class, and record the cloud region that serves it. If the vendor does not say where, write down the assumption and who made it.
    4. Apply the factors. Multiply tokens by the factor for the class and region, in kilograms of CO2e per million tokens. For message counts, convert to tokens with stated defaults first.
    5. Aggregate. Add the lines by service, team and month. Keep each line separate until the very end so the data tier of each stays visible.
    6. Document assumptions and uncertainty. Record the factor version, the class and region for each model, every default used, and the range. An estimate with its range is more useful to an auditor than a single number.
    7. Report under Scope 3. Purchased AI services belong in Scope 3 Category 1, purchased goods and services, of the GHG Protocol. State the method tier next to the figure and keep the evidence.

    Worked example

    A company has one API workload of 100 million tokens a year on a mid-size model (class B) served from US East - N. Virginia, and 25 active seats of a chat assistant, assumed to run on the same class and region.

    Estimated kg CO2e for one year. The seat line uses the dataset's default messages per active seat. Estimates.
    LineDataCentralRange
    API workloadTier 2a, billed tokens2.61up to 39% either way
    Chat seatsTier 2b, 9,000 messages0.5020.25 to 1.2
    TotalMixed tiers3.11shown per line

    Report the two lines separately. Adding them is fine for a headline, but a reader should see that one rests on exact token counts and the other on a benchmark for messages per seat. Served from EU North - Sweden, the API line would be 0.714 kg, so the region assumption matters more than most other choices.

    Common mistakes

    • Counting only the chatbot staff talk to and missing API calls, coding assistants and AI inside other software.
    • Using spend for everything. It is the lowest quality tier and the least useful for finding what to change.
    • Leaving the region and model class unstated, so nobody can rebuild the number.

    Go deeper on each step

    Keep reading