SOMA

    What does the GHG Protocol say about reporting emissions from AI and data center services?

    For sustainability teams placing AI and cloud services in a GHG inventory. A paraphrase and not the text of the standards, so read the linked documents before relying on it. Checked on 12 November 2026. Not legal or assurance advice.

    Page updated

    What does the GHG Protocol say about reporting emissions from AI and data center services?

    It has no AI-specific rule, so the general rules apply. AI and cloud services you buy are upstream emissions of purchased services and go in Scope 3 Category 1. A data centre you own or control is reported in Scope 1 and Scope 2 according to your consolidation approach, with the electricity in Scope 2. Hardware you buy is capital goods in Category 2. The Protocol offers four methods to calculate a purchased-service line, from supplier-specific data to spend. We did not find AI or data centres named on the Scope 3 Technical Guidance index page we read, so the placement of AI services is an application of the general categories and not a quotation.

    Where does each kind of AI or data-centre activity fall?

    Our reading of the Scope 3 Standard and the Corporate Standard. Boundary choices change the answer, so confirm the case with your assurance provider.
    ActivityWhere it fallsWhy
    Paying for AI APIs, enterprise AI seats or cloud servicesScope 3, Category 1 (purchased goods and services)The supplier runs the hardware, and the cradle-to-gate emissions of a purchased service are Category 1
    A data centre you own or operate, buying electricityScope 2 for the electricity, Scope 1 for on-site fuel and refrigerantsWithin your organisational boundary. Report Scope 2 with both the location-based and market-based methods
    Space or racks leased from a data-centre or colocation operatorScope 1 and 2 if your consolidation approach puts the leased space inside your boundary, for example operational control over it. If not, Scope 3, Category 8 (upstream leased assets)The consolidation approach decides, so state which one you use
    Servers and accelerators you buyScope 3, Category 2 (capital goods)Cradle-to-gate emissions of purchased equipment
    Training a model on rented cloud capacityScope 3, Category 1It is a purchased cloud service. Training your own model on your own hardware is Scope 1 and 2

    The same distinction explains the line in the CSRD standards between cloud services and the rest of Category 1. See how to report AI under ESRS E1.

    What are the four Scope 3 calculation methods, and which fits AI?

    The Scope 3 Technical Guidance sets out four methods for purchased goods and services, in the order from most to least specific to the supplier.

    • Supplier-specific. The supplier gives emissions for your purchase. We found no AI API or seat vendor offering this with a stated boundary today. Cloud provider dashboards report emissions for your own cloud account.
    • Hybrid. The supplier gives energy or partial data, and secondary factors fill the gaps. A provider's energy report for your account would fit here.
    • Average-data. Physical activity, such as billed tokens, multiplied by a published factor. This is how SOMA's Tier 2a and Tier 2b work.
    • Spend-based. Money spent multiplied by an economic intensity factor. SOMA's Tier 1 uses 0.09613 kg CO₂e per euro with an uncertainty of ±300%.

    The guidance does not require the most specific method first. It asks for the best data you can obtain and a note of the method. Spend-based figures for the same purchase ran 1.7 to 12.3 times a token-based estimate in SOMA's method paper, which was accepted for publication in Sustainable Production and Consumption, so a spend line is a screening step.

    What is still changing?

    The GHG Protocol standards are under revision, and parts of the proposed text are still drafts. The site tracks which parts are final and how the data tiers map to the draft data types in a separate guide. Until the revision is final, document your method, the factor version and the uncertainty next to each line. The method behind SOMA's factors is on the methodology page, and the step-by-step inventory is in the Scope 3 guide. SOMA AI publishes this page and sells AI footprint reports, so check the Protocol documents yourself.

    Keep reading