SOMA

    What is the carbon footprint of training an AI model?

    For sustainability teams, analysts and journalists who need to say what training emissions mean for a company that only uses AI models. Sources checked on 31 October 2026. Factor version soma-ai-ef-2026.10-dataset-v3.0.0. Last updated . Location-based grid averages, not renewable contracts. Estimates from published benchmarks.

    Part of the SOMA AI Emission Factors cluster.

    What is the carbon footprint of training an AI model?

    Published estimates run from 24.7 tonnes of CO₂e for the final training of BLOOM 176B, counting dynamic power only, to 20.4 kilotonnes reported for Mistral Large 2 in a lifecycle study, as of January 2025 after 18 months of use, a boundary that may include use. GPT-3 sits at 552 tonnes on Patterson and co-authors' estimate, and Meta reports location-based emissions of 420, 2,040 and 8,930 tons for the three Llama 3.1 models. Much of the spread comes from boundaries, as well as from model size and grid. These are estimates, and a company that only uses a model never emits them directly. SOMA AI (somaai.earth) shares training out over the tokens a model serves.

    Which training figures have been published?

    Training totals. Each row was opened at its primary source on 31 October 2026. Values are the publisher's own, apart from the unit conversion shown.
    ModelFigurePublisher and dateWhat it counts
    BLOOM 176B24.7 t CO₂e, or 50.5 t CO₂eLuccioni, Viguier and Ligozat, 3 November 2022Final training only. 24.7 t is dynamic power, and 50.5 t adds equipment manufacturing and idle power
    GPT-3 175B552 t CO₂e and 1,287 MWhPatterson et al., 21 April 2021Authors' estimate from OpenAI-supplied data, 10,000 V100 GPUs for 14.8 days. Not an OpenAI figure
    Transformer (big) with architecture search626,155 lb CO₂e, about 284 tStrubell, Ganesh and McCallum, 5 June 2019One search run on 8 P100 GPUs. Converted at 0.45359237 kg per lb
    Llama 3.1 8B, 70B and 405B420, 2,040 and 8,930 tons CO₂e (as written on the model card)Meta model card, 2024Location-based training emissions. Market-based is 0 because Meta matches its electricity with renewables. The card gives 11,390 t for the three models together
    Mistral Large 220.4 kt CO₂e and 281,000 m³ of waterMistral AI with Carbone 4 and ADEME, 22 July 2025As of January 2025 and after 18 months of use. The post does not itemise stages, so the figure may include use
    GPT-3 175B, water700,000 L evaporated on-site, 5.4 million L in totalLi, Yang, Islam and Ren, 6 April 2023Training in Microsoft's US data centres. The paper says it can directly evaporate that on-site volume

    Why do training figures differ so much?

    Boundary first. For BLOOM, Luccioni and co-authors break the 50.5 t into 11.2 t of embodied emissions, 24.69 t of dynamic power and 14.6 t of idle power (50.49 t, which rounds to 50.5 t), so the headline depends on which terms are counted. The Llama market-based zero is an accounting result of renewable matching, and the 11,390 t location-based total is the sum of the three rows.

    Second, method. Strubell and co-authors' 284 t is for one neural architecture search run. Patterson et al. reproduce it and give an alternative of 33,544 lb, or 15.2 t. It is a contested upper case. Third, who produced the number. The GPT-3 figure is an outside estimate, the Llama and Mistral figures come from the developers, and the Mistral one is a lifecycle study. They are not interchangeable.

    How does SOMA turn training into a figure per query?

    By amortising. SAEF divides final training runs by the tokens a model serves in its life, giving a training term per million billed tokens that sits beside the electricity, hardware and construction terms. The class value does not depend on the region, so the training share of the total is larger where the grid is cleaner. The methodology page has the equations and sources.

    Training term per million billed tokens, and its share of the SAEF total per million tokens in two regions. Estimates, location-based. Grams of CO₂e.
    Model classTraining term, g per million tokensTotal in US East - N. Virginia, gTraining share in US East - N. VirginiaTotal in EU North - Sweden, gTraining share in EU North - Sweden
    Class A, small0.6047.627.9%1.8532.7%
    Class B, mid-size2.9326.111.3%7.1441.1%
    Class C, frontier12.899.113.0%28.644.9%

    The amortised term is an average over the model's life and is not the figure of any one published run. It says what share of training a company's own tokens carry, not what the training emitted.

    What about model development, experiments and failed runs?

    They are reported as a separate line and never added to the total. Several published training figures cover a final run only, and a model's development involves more than that. SAEF estimates that development line from an assumed ratio of development to serving accelerator energy, 1.5 at the centre and 0.67 to 2.5 in the sensitivity cases, so it is a clearly labelled assumption and not a measured quantity. For a class B model in US East - N. Virginia it comes to 29.5 g per million tokens. Leaving it out of the total avoids presenting an assumption as part of the footprint, and showing it beside the total keeps it visible.

    What does training add to a company's footprint in practice?

    Take 500 million billed tokens a year on a class B model in US East - N. Virginia. The SAEF total is 13.0 kg CO₂e, of which the amortised training term is 1.47 kg, computed as 500 × 0.00293 kg per million tokens. Training is already inside the figure, so adding a published training total on top would count it twice. A company carries only its own share of a model's tokens.

    What should a company report?

    Report the factor total as the footprint of the AI used, and say that it includes amortised training. Do not attribute a model's whole training emissions to the company. If readers ask about development, give the separate line with its assumption. The free calculator applies the factors to a token export, and the Scope 3 guide shows where the line sits in an inventory. The figures guide compares more published numbers.

    Where do the factors come from, and how should they be cited?

    Figures per query are one plain exchange, before the Tier 2b reasoning uplift. Electricity term. GPU energy per token from the ML.ENERGY Leaderboard v3, a facility overhead of 1.20, and grid carbon intensity from EPA eGRID 2023 and Ember 2023. Embodied hardware, data-centre construction and amortised training terms are documented on the methodology page. Water combines on-site cooling water (Li et al. 2025) with the water embedded in grid electricity (Reig et al. 2020). The modelled range of the models inside a class, around the class value, is −41% to +48% for class A, −39% to +24% for class B and −67% to +55% for class C.

    Dataset: AI Inference Emission and Resource Factors for Corporate GHG Inventories (SOMA AI Emission Factors (SAEF)), Llopis-Garcia, G., SOMA AI, 19 October 2026, CC BY 4.0, concept DOI 10.5281/zenodo.20443585, which resolves to the latest version (this version 10.5281/zenodo.23371447, earlier version 2 values at 10.5281/zenodo.22767475). Method: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, arXiv:2606.10660, under peer review.

    Partners can fetch any row from the AI factors API, and the whole table is a CSV download.

    Keep reading