SOMA

    Is AI bad for the environment?

    For sustainability teams, students and journalists who want the evidence at both scales. 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.

    Is AI bad for the environment?

    One query is small and the total is growing fast. The International Energy Agency puts all the world's data centres, not only AI, at 415 TWh in 2024, 1.5% of world electricity, and projects around 945 TWh by 2030. A single text prompt, by contrast, is estimated at 0.24 Wh by Google and about 0.3 Wh by Epoch AI. Both statements hold at once. Whether AI is bad for the environment depends on the volume, the grid and the model, and the evidence below lets a reader judge each. All figures are estimates or projections unless the row says otherwise.

    How much electricity do data centres use in total?

    The totals come from a US national laboratory and the International Energy Agency, plus two scenarios by one researcher, and include every use of a data centre, so they overstate what AI alone uses. The projections are scenarios, not forecasts.

    Data-centre and AI electricity, as published.
    Publisher and dateFigureScopeWhat it is
    Lawrence Berkeley National Laboratory, 20 December 2024176 TWh in 2023US data centres, 4.4% of US electricityAll uses, not AI only. Up from 58 TWh in 2014.
    Lawrence Berkeley National Laboratory, 20 December 2024325 to 580 TWh in 2028US data centres, 6.7% to 12% of US electricityProjection, a scenario range, all uses.
    International Energy Agency, April 2025415 TWh in 2024, around 945 TWh in 2030World data centres, 1.5% of world electricity in 2024All data centres. 945 TWh is a projection.
    de Vries, Joule, 18 October 202385 to 134 TWh a year by 2027About 1.5 million NVIDIA AI serversScenario, if all run at full power. Narrower than AI worldwide.
    de Vries, Joule, 18 October 202329.2 TWh a yearHypothetical, every Google search using AIScenario from SemiAnalysis's server count, not a measurement.

    Two cautions apply. The US and world figures are not AI-only, so they cannot be read as the cost of chatbots. And the 2023 scenarios by de Vries were built before the 2025 per-query estimates and should be read as such.

    How big is one query against these totals?

    Very small. Google reports 0.24 Wh and 0.03 g CO2e for its median Gemini Apps text prompt, counting chips, host machines, idle capacity and data-centre overhead, with a market-based emission factor and training excluded. Epoch AI modelled about 0.3 Wh for GPT-4o with 500 output tokens. The scale changes with long inputs and reasoning, where Epoch AI estimates almost 40 Wh for a 100,000-token input and Microsoft researchers a median 3.91 Wh for test-time-scaling queries. The per-query detail is on the ChatGPT energy page.

    What about water?

    Google reports consuming 10.9 billion US gallons of water in 2025 across its data centres and offices, about 41 billion litres, and the figure covers the whole company, not only AI. Li and co-authors project global AI demand at 4.2 to 6.6 billion cubic metres of water withdrawal in 2027, a projection of withdrawal and not of consumption. Per query, the same group estimated 16.904 mL for a medium-length GPT-3 response at the US average, counting both on-site cooling and the water used to generate the electricity. What each water figure counts is set out in the water guide.

    What is not known?

    • How much of the total data-centre electricity is AI. The agency totals count every use.
    • What most commercial services use. Google and Mistral publish methods, while the 0.34 Wh per ChatGPT query that OpenAI's chief executive Sam Altman gave comes with none, and the rest of the figures are researchers' models or open-model benchmarks.
    • How fast per-query energy is falling against how fast volume is rising. Both are moving, and the 2025 estimates for a short query are about a tenth of the 2024 EPRI figure of 2.9 Wh.
    • The emissions of making the chips and servers, which many per-query figures leave out (Google, Mistral and Berget AI include them).

    What can a company using AI do about it?

    Two choices shape the footprint more than any other, and both are in the company's hands. The first is the region. With SOMA AI (somaai.earth) factors, one exchange on a class B model comes to 0.0094 g CO2e when served from EU North - Sweden and 0.055 g from Asia Pacific - Singapore, a factor of 5.9. Where a provider lets the customer pick a region, that is a direct lever. A lower-carbon grid can raise water use, so check both.

    The second is the model class. The same exchange is estimated at 0.03 Wh on a small class A model and 0.37 Wh on a large class C model, a factor of 12. Using the smallest model that does the job follows from that. A third lever is not to send long inputs or reasoning modes where a short query would do.

    What should a company report?

    Report the AI footprint as an estimate with its basis, in the Scope 3 inventory and beside it the water. Name the activity data, the model class, the region and the dataset version. 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.

    Keep reading