SOMA

    How does the carbon footprint of AI compare with driving, flying, streaming and food?

    For journalists, analysts and sustainability teams who want a number set against something familiar, with the boundary of each. Sources checked on 10 November 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.

    How does the carbon footprint of AI compare with driving, flying, streaming and food?

    One mid-size AI exchange of 1,210 tokens is estimated by SOMA AI (somaai.earth) at 0.053 g CO₂e on the world-average grid, about as much as 32 centimetres in an average UK car or 5 seconds of streaming video. A 100-seat company using AI chat for a year comes to an estimated 2 kg CO₂e on SOMA's Tier 2b defaults (class B in AWS us-east-1, reasoning uplift included), about 12 km of driving. Each comparison below carries the boundary of both sides, because a familiar number only helps when it counts the same things.

    How does one AI query compare with everyday things?

    The last column divides each figure by the SOMA estimate of 0.053 g per exchange, a class B model at the world-average grid, location-based. Published AI figures differ by their own boundaries. Google reports 0.03 g CO₂e for a median Gemini Apps text prompt on a market-based factor with training excluded. Mistral reports 1.14 g for a 400-token Le Chat response, counting upstream emissions such as manufacturing. Berget AI, in a company release, reports 0.4 g and 3.2 g for a Kimi K3 request of 5,000 input and 387 output tokens in Sweden and Texas. It is location-based and includes hardware manufacturing, with training excluded.

    Everyday figures, each with what it counts, against the number of SOMA-estimated exchanges that equal it.
    ActivityFigure and boundarySourceEquivalent AI exchanges
    One Google search0.2 g CO₂, server side, 2009, before AI featuresGoogle, 20093.8
    One smartphone charge0.019 kWh, so 19 Wh. Compared on energy because the EPA carbon figure uses the US gridUS EPA190 (at 0.099 Wh each)
    One hour of video streaming36 g CO₂, 2019, operational electricity. Of the energy, devices 72%, networks 23%, data centres 5%IEA, 2020680
    One km, average UK car165.91 g CO₂e per vehicle-km, tailpipe only, no fuel production or manufactureUK DESNZ, 20263,100
    One km, average new EU car96.7 g CO₂, type-approval, new cars registered in 2025 only, electric cars counted at zero, provisionalEEA, 20261,800
    One passenger-km, short-haul flight127.86 g CO₂e with radiative forcing, 75.59 g without, UK flights, 8% distance upliftUK DESNZ, 20262,400 with, 1,400 without
    100 g of beef99.48 kg CO₂e per kg, full supply chain including land use and losses, global averageOur World in Data, from Poore and Nemecek 2018190,000
    100 g of poultry9.87 kg CO₂e per kg, same boundaryOur World in Data, from Poore and Nemecek 201819,000

    The equivalents are the figure in grams divided by 0.053, rounded to two significant figures. The car, flight and streaming rows mix places and years, so read them as orders of magnitude.

    How does a company's annual AI use compare with driving or flying?

    A year is the fairer unit for a company, because a single query is too small to feel. For 100 seats using chat on a class B model served from AWS us-east-1, the SOMA Tier 2b defaults give 36,000 exchanges, including the default reasoning uplift, and an estimated 2 kg CO₂e, location-based. Divided by the UK average car factor that is 12 km of driving, and divided by the short-haul with-radiative-forcing factor it is 16 passenger-km of flying. Both comparisons put the company's yearly chat use at the scale of one short car journey, not of a fleet. The uplift multiplies the output tokens of the average exchange by 2.2, because some chat messages use reasoning, as explained in how much more energy reasoning models use. Heavier use changes this. Agentic coding and document work send far more tokens per exchange, and the estimate then rises in proportion.

    How do AI and data centres compare with aviation at global scale?

    The International Energy Agency puts all the world's data centres, not only AI, at about 180 Mt of indirect CO₂ today from the electricity they use, excluding backup generation, and 415 TWh of electricity in 2024, 1.5% of world electricity. The International Civil Aviation Organization puts aviation at 2.5% of global energy-related CO₂ in 2023 and, historically, 2.4% of anthropogenic CO₂ and about 4% of radiative forcing. The IEA gives the data-centre figure as 0.5% of combustion emissions.

    These are not the same metric. The data-centre number is indirect emissions from the electricity bought, the aviation number is fuel burned in flight, and the 415 TWh is electricity while aviation uses jet fuel. The two shares sit on similar but not identical denominators, so they show the order of magnitude and not an exact ratio. Neither number isolates AI, which is a subset of data-centre load.

    How does the water per query compare with beef, a shower or a flush?

    SOMA estimates 0.24 mL of water per exchange on a class B model in AWS us-east-1, counting on-site cooling and the water behind the electricity. Google reports 0.26 mL for a median Gemini Apps prompt, from data-centre water at its fleet WUE of 1.15 L per kWh. It does not say whether the water behind the electricity is included. The household figures are US EPA fixture standards, not UK or global averages, and the beef figure counts mostly rainwater.

    Water figures with their basis, against one SOMA-estimated exchange.
    ItemFigureSourceEquivalent AI exchanges
    One toilet flush1.6 US gallons, 6.1 L, US federal standard (WaterSense 1.28 gal)US EPA WaterSense25,000
    One minute of showering2.5 US gallons per minute, 9.5 L, US standard showerheadUS EPA WaterSense40,000
    100 g of beef, total water footprint15,400 L per kg, so 1,540 L, 94% rainwaterMekonnen and Hoekstra 20106,500,000
    100 g of beef, blue water only550 L per kg, so 55 L, surface and groundwaterSame report230,000

    Only the blue-water row is close in kind to cooling water. Setting 15,400 L of mostly rain against millilitres of cooling water overstates the contrast.

    Why do these comparisons mislead?

    • Tailpipe against lifecycle. The UK car factor counts only what leaves the exhaust. SOMA's AI estimate includes hardware, construction and amortised training, so it is the wider of the two.
    • Location against market. SOMA reports location-based emissions on the grid where the data centre sits. Google's 0.03 g uses a market-based factor of 94 g/kWh, which reflects its contracts.
    • Per query against per year. A query is tiny, a year of company use is not, and the sector total is larger again. Pick the unit that matches the question.
    • Reasoning multiplies the figure. The class B estimate is for a plain exchange. SOMA's Tier 2b rule applies a central multiplier of 25 to the output tokens of reasoning exchanges. The detail is in how much more energy reasoning models use.
    • Years and places differ. The car and flight factors are UK 2026, the streaming figure is a 2019 global average and the Google search figure is from 2009, while the SOMA figure uses the current world-average grid.
    • Devices in or out. The streaming figure includes the viewer's screen and network, which are most of it. SOMA's AI estimate covers the data centre only, not the user's device or network.
    • One caveat on the figures. Sam Altman's 0.34 Wh per ChatGPT query is from his personal blog and states no boundary, so it cannot be placed on this table.

    What should a company report?

    Report the AI footprint as an estimate in the Scope 3 inventory, with the model class, region, activity data and dataset version beside it, and keep the location-based figure as the main line. Comparisons such as these belong in commentary, not in the inventory. The method is in the Scope 3 guide, the equations are on the methodology page, and the free calculator gives a first estimate without an account.

    Where do the figures come from?

    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