SOMA

    Why is Google's 0.24 Wh per Gemini prompt figure so different from academic estimates?

    For analysts, journalists and sustainability teams comparing Google's published figure with other estimates. Sources checked on 12 November 2026. SOMA figures are estimates from the open SAEF dataset.

    Page updated

    Why is Google's 0.24 Wh per Gemini prompt figure so different from academic estimates?

    It differs less than the headline suggests. For a short text prompt, the academic and third-party estimates of 0.2 to 0.4 Wh sit close to Google's 0.24 Wh. The gap appears against older figures such as 2.9 Wh, against measurements of chips alone, which come out lower, and against estimates for long or reasoning prompts, which come out higher. Google measured its own production fleet for the median prompt, and academic work models other services or runs benchmarks, so the boundary and the prompt differ.

    What did Google measure?

    Google's paper (Elsworth et al., arXiv 2508.15734, 21 August 2025) reports the median Gemini Apps text prompt in May 2025. The stated boundary is the full stack of serving infrastructure, which is the active AI accelerator power, host system energy, idle machine capacity and data-centre energy overhead. Training is excluded. The paper and the companion blog post give 0.24 Wh, 0.03 g CO₂e and 0.26 mL of water. The carbon figure uses the market-based method with Google's 2024 fleet-wide carbon intensity, 94 g per kWh, and includes the embodied emissions of the AI hardware. Google also reports a fleet-wide PUE of 1.09 and a 33 times fall in energy and 44 times fall in carbon per median prompt over the twelve months to May 2025.

    The same blog gives a narrower number on the same prompt. Counting only active TPU and GPU consumption, it is 0.10 Wh, 0.02 g CO₂e and 0.12 mL, which Google calls an optimistic scenario.

    How does it compare with academic and third-party estimates?

    Energy for one short text query, as published, and the SOMA estimate. Boundaries differ, so read the last column.
    Source and dateFigureQueryBoundary
    Google (Elsworth et al.), August 20250.24 WhMedian Gemini Apps text prompt, measuredFull serving stack including idle machines and overhead.
    Google, August 20250.10 WhSame promptActive TPU and GPU only.
    Epoch AI, February 20250.3 WhGPT-4o, 500 output tokens, modelledIncludes server and data-centre overhead.
    Oviedo et al. (Microsoft), September 20250.31 Wh (IQR 0.16 to 0.60)Models above 200 billion parameters, 300 median output tokens, modelledH100 nodes, with data-centre overhead.
    Jegham et al., May 20250.42 Wh (0.37 Wh without overhead)GPT-4o, 100 input and 300 output tokens, modelled from API latencyEstimated hardware, with overhead.
    ML.ENERGY Leaderboard, February 20260.21 WhLlama 3.1 405B, 367 output tokens, measuredGPU energy only, 8 H100 GPUs.
    SOMA AI estimate0.03, 0.099 and 0.37 Wh for classes A, B and C928 tokens in, 282 out (1210 in all)Facility level, with host, idle and PUE terms.

    Why do the numbers still differ?

    • Median against mean. Google reports the median prompt, which is short. Long prompts and reasoning raise the mean well above the median, and academic estimates often fix a token count that is larger than the median.
    • Measured against modelled. Google read its production fleet. Academic and third-party work infers energy from benchmarks, API latency and assumed hardware, usually on GPUs and not on Google's TPUs.
    • Boundary. Chips-only figures such as the ML.ENERGY 0.21 Wh and Google's own 0.10 Wh leave out the host, idle capacity and overhead. SOMA multiplies GPU energy by 1.50 for the host server and 1.10 for idle capacity, and applies a PUE of 1.20 against Google's 1.09.
    • Efficiency of one provider. The 33 times fall is Google's own result on its own service. It does not transfer to other providers.
    • Carbon accounting. The 0.03 g uses a market-based factor of 94 g per kWh. A location-based grid factor gives a different number for the same energy, so carbon figures from Google and from SAEF are not interchangeable.
    • The prompt. SOMA's exchange is 1210 tokens, longer than a median chat prompt, which is why its class estimates straddle Google's figure.

    Google's 0.24 Wh falls between the SOMA class B and class C estimates for a 1210-token exchange, and it is not a statement about any other provider. These are estimates, not meter readings from the other services. The method paper was accepted for publication in Sustainable Production and Consumption in November 2026, and the terms are on the methodology page.

    Keep reading