P.03AI Footprint

    AI works for you. Know what it costs the planet.

    AI is now part of your footprint. SOMA estimates the energy, emissions and water behind your AI workloads and folds them into the same reporting as everything else.

    01

    Account for AI use

    Estimate the energy, emissions and water of the models and tools your teams actually run, per workload and per provider.

    • Per tool and per team, per model where token exports exist
    • Energy, emissions and water metrics
    • Based on the usage data you have, from token exports down to licence counts, and published, open emission factors
    02

    Report it properly

    AI impact lands in your existing inventory, categorised and evidenced like any other emission source.

    • Purchased AI services in Scope 3 Category 1
    • Same factor traceability as everything else
    • Documented for customer questionnaires, assurance and voluntary or CSRD reporting
    03

    Reduce with intent

    See which workloads drive the impact and where smaller models or greener regions cut it without losing capability.

    • Hotspot analysis across teams
    • Model and region comparisons
    • Reduction levers in your decarb plan

    How much CO2e does AI usage emit?

    A mid-size AI model emits about 0.039 kg CO2e per million billed tokens on the world-average electricity grid. Small models are about 0.012 kg and frontier models about 0.147 kg. Region changes the figure several times over: the same mid-size model is about 0.007 kg in Sweden and 0.040 kg in Singapore.

    Small model

    0.012

    Mid-size model

    0.039

    Frontier model

    0.147

    kg CO2e per million billed tokens · world-average grid

    Source: SOMA AI Emission Factors, dataset v3.1.0 (12 November 2026), CC BY 4.0. Modelled estimates, not measurements of any provider's infrastructure.

    See the full factor table

    Models and providers covered

    The SOMA AI Emission Factors cover more than 150 text model families from 23 providers, grouped into three size classes. Any model id that is not listed is assigned the frontier class, so no usage is left out.

    OpenAI
    GPT-6, GPT-5, GPT-4.1, GPT-4o, o3, o4-mini, Codex, gpt-oss
    Anthropic
    Claude Opus, Claude Sonnet, Claude Haiku
    Google
    Gemini Pro, Gemini Flash, Gemma
    Mistral
    Mistral Large, Medium and Small, Magistral, Codestral, Devstral, Ministral, Pixtral
    Meta
    Llama 4, Llama 3
    Alibaba
    Qwen
    DeepSeek
    DeepSeek V4, V3, R1
    Amazon
    Nova, Titan
    Cohere
    Command, Aya
    xAI
    Grok
    Microsoft
    Phi
    NVIDIA
    Nemotron
    Perplexity
    Sonar
    Moonshot AI
    Kimi
    Z.ai
    GLM
    Also
    AI21 Jamba, MiniMax, Tencent Hunyuan, Xiaomi MiMo, Writer Palmyra and others

    Image, video, speech and embedding models are out of scope. Model names are trademarks of their owners and do not imply endorsement.

    Cloud regions covered

    143 cloud regions on AWS, Google Cloud and Azure, plus any country and every US eGRID subregion. Each region uses its own grid electricity intensity.

    Built on open data

    The factors combine published GPU energy benchmarks with grid electricity data and hardware lifecycle terms. Sources: ML.ENERGY Leaderboard, Ember yearly electricity data, US EPA eGRID, and published water-use studies. The method is described in a preprint, accepted for publication in Sustainable Production and Consumption.

    Last updated: 12 November 2026.

    Questions about
    AI emissions.

    Need the AI line in your inventory documented for audit?

    AI footprint report for your organisation from €3,000 (excl. VAT), fixed price, scoped on a short call.

    Use AI for the planet. Start by accounting for it.