SOMA

    How much energy does DeepSeek use, and is it more efficient than ChatGPT?

    For sustainability teams, analysts and journalists who need a defensible figure for DeepSeek. 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.

    How much energy does DeepSeek use, and is it more efficient than ChatGPT?

    Published modelling puts DeepSeek-R1 at 29.078 Wh for one long query of 10,000 input and 1,500 output tokens on DeepSeek's own servers, the highest of the 27 models in the study, and 7.41 Wh when the same model runs on Azure. We found no public per-query comparison of DeepSeek and ChatGPT on the same basis, so we cannot say which is more efficient. SOMA AI (somaai.earth) estimates a DeepSeek-R1 exchange of 1,210 tokens at 0.37 Wh, the same as any model in its largest class.

    What do the published measurements say about DeepSeek models?

    Two sources give DeepSeek numbers we could verify, and they count different things. Read the boundary before the figure.

    Published energy figures involving DeepSeek models, with what each counts.
    SourceFigureModel and queryWhat it counts
    Jegham et al., May 202529.078 Wh (plus or minus 9.725). 7.41 Wh on AzureDeepSeek-R1, 10,000 input and 1,500 output tokens. The study range for 27 models is 0.443 to 29.078 WhModelled from public API latency and estimated hardware, with data-centre overhead. Inference only.
    AI Energy Score v2 (Hugging Face and partners), December 20250.050 Wh per query with reasoning off, 7.63 Wh onDeepSeek-R1-Distill-Llama-70B, a distilled open model and not DeepSeek-R1 itselfGPU energy only, as the column is labelled, measured with CodeCarbon. 49.53 and 7,626.53 Wh per 1,000 queries.

    The Jegham figure is a long, reasoning-heavy query on one provider's own infrastructure, so it is a high-end case and not a typical one. A 2025 study by Dauner and Socher also measured DeepSeek-R1 70B on an A100 GPU, but its published range joins two question phases, so we leave it out.

    Why does reasoning dominate DeepSeek's numbers?

    Because a reasoning model writes many more tokens before it answers. For the distilled 70B model, the AI Energy Score source gives 154 times more GPU energy with reasoning on than off, 7.63 Wh against 0.050 Wh per query. The same source reports that reasoning models use, on average, 30 times more energy than models without it. The same model also moves by a factor of 3.9 between DeepSeek's servers and Azure in the Jegham modelling, so where a model is hosted also moves the figure. See how much more energy reasoning models use.

    How does SAEF class DeepSeek models, and where does it assume they run?

    SOMA does not know DeepSeek's hardware or serving region. It classes deepseek-r1 and deepseek-v3 as the largest class from their published mixture-of-experts size, with 671 billion parameters in total and 37 billion active, and the distilled models as mid-size or small by parameter count. SAEF assigns no serving region for DeepSeek, so the table uses the world-average grid. If you run a DeepSeek model in a cloud region of your own, use that region instead.

    Estimated electricity and carbon for one exchange of 1,210 tokens on the world-average grid. Estimates. The GPT rows are the same class factors, shown for comparison.
    ModelSAEF classWhg CO₂e
    DeepSeek-R1, reasoning by defaultClass C0.370.2
    DeepSeek-V3 and V3.1Class C0.370.2
    R1-Distill-Llama-70B and R1-Distill-Qwen-32BClass B0.0990.053
    R1-Distill-Qwen-14B, R1-Distill-Llama-8B and smallerClass A0.030.016
    GPT-5, for comparisonClass C0.370.2
    GPT-4o, for comparisonClass B0.0990.053
    GPT-5 mini, for comparisonClass A0.030.016

    For the long Jegham query, the SAEF class estimate is 2.9 Wh, far below the 29.078 Wh modelled for DeepSeek's own servers. The methods differ and SAEF applies no reasoning multiplier to a fixed token count, so a reasoning model with a long hidden chain would sit above this estimate.

    Is DeepSeek more efficient than ChatGPT?

    We cannot answer that from the sources we checked. The Jegham data places DeepSeek-R1 at the high end, but the rows we verified carry no GPT figure on the same basis. At the same SAEF class the two return the same estimate, which reflects the method and not a measured tie. Claims about training cost sit outside our source check and we do not repeat them. We found no published per-query figure from DeepSeek in our source check.

    What should a company report?

    Report the model, the class, the region you assume, the token source and whether reasoning tokens are included in the count. Where DeepSeek is used through a hosted API, take tokens from the provider's usage records and check whether its billed output count includes the reasoning tokens. SAEF assumes it does for reasoning models, and applies no extra multiplier. The method is in the Scope 3 guide and on the methodology page, and why published figures differ sets the boundaries side by side. The free calculator gives a first estimate.

    Keep reading