> ## Content Index > Fetch the complete content index at: https://unhyd.com/llms.txt > Use this file to discover other available public pages before exploring further. # How Much Electricity Does an AI Query Use? Why One Number Misleads - URL: https://unhyd.com/article/ai-query-electricity-use-per-prompt-explained/ - Published: 2026-10-20T21:28:03.000Z - Updated: 2026-10-20T21:28:03.000Z - Description: A 0.24 Wh Gemini median and a 0.34 Wh ChatGPT claim cannot settle AI's energy debate. Here is what a careful comparison actually shows. - Author: Unhyd Editorial - Tags: AI, Artificial Intelligence, AI Infrastructure, Energy Consumption, Data Centers, Sustainability **AI query electricity use** has become a shorthand for the environmental cost of artificial intelligence. One widely cited figure is 0.24 watt-hours for a median text prompt in Google's Gemini Apps; another is 0.34 watt-hours for an average ChatGPT query, stated by OpenAI CEO Sam Altman. Neither figure answers how much electricity all AI uses, and they are not directly comparable measurements of identical workloads. That distinction is more than statistical caution. The first number comes from a Google-authored technical paper describing a production measurement boundary. The second appeared in a 2025 personal essay without a published supporting measurement method. Meanwhile, the [International Energy Agency's April 2026 analysis](https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary?ref=unhyd.com) says electricity use at AI-focused data centers grew by 50% in 2025\. Small reported figures for individual text requests and rapidly growing aggregate demand can both be true. ## What the AI query electricity use figures actually measure In a [technical paper published in August 2025](https://arxiv.org/abs/2508.15734?ref=unhyd.com), Google researchers reported that the median Gemini Apps text prompt consumed 0.24 Wh. The authors said their approach counts not only the active AI accelerator but also host computing, idle capacity kept ready for traffic and data-center overhead. Their measurement concerns Gemini Apps text prompts in a specified period; it is not a reading for every model, every user or every type of generation. In his [June 2025 essay](https://blog.samaltman.com/the-gentle-singularity?ref=unhyd.com), Sam Altman wrote that an average ChatGPT query used about 0.34 Wh. The statement is a first-party claim about a product average, not a detailed public audit. Google's own paper notes that Altman's figure did not explain the measurement boundary. It would therefore be misleading to declare one company's assistant more energy-efficient than another simply because 0.24 is smaller than 0.34. There is a further mathematical mismatch: a *median* is the middle observation in a distribution, while a *mean* is the total divided by the number of observations. If some tasks are unusually demanding, the mean can be much higher than the median. A published median cannot be multiplied by the actual number of prompts to establish a fleet's real total unless the analyst knows the rest of the distribution. ## An original calculation: what happens when the same assumption is scaled To make the units tangible, Unhyd calculated two purely hypothetical scenarios using the two published per-prompt figures as **fixed illustrative assumptions**. This is arithmetic, not a measurement of anyone's AI usage, and it does not imply that the 0.24 Wh Gemini median is an observed average. In particular, the 0.34 Wh value is not a controlled comparison against the Google measurement. | Illustrative activity | At 0.24 Wh per prompt | At 0.34 Wh per prompt | | ------------------------------------- | --------------------- | --------------------- | | 100 prompts in one day | 24 Wh | 34 Wh | | 100 prompts daily for 365 days | 8.76 kWh | 12.41 kWh | | 10 billion prompts daily for 365 days | 0.876 TWh | 1.241 TWh | Illustrative calculations using fixed per-prompt assumptions. These are not measured consumption totals or direct product comparisons. The formula is straightforward: number of prompts × assumed watt-hours per prompt ÷ 1,000 gives kilowatt-hours. For the large-scale example, dividing the annual watt-hour result by one trillion gives terawatt-hours. The 10-billion-prompts-per-day volume is a deliberately round hypothetical, **not** a verified count of real daily AI requests. Notice what the calculation does and does not establish. It shows how a small unit value becomes meaningful at enormous scale. It does not establish total AI electricity use, because AI workloads are heterogeneous, model providers do not publish a complete reconciled request census, and data-center electricity includes activities beyond responding to consumer text prompts. ![Three-stage decision diagram showing how to confirm a real job opening, verify the recruiter's authority and check the onboarding process before sharing sensitive information.](https://storage.ghost.io/c/df/e4/dfe4d080-643c-453a-8a60-0452bdbc0efd/content/images/2026/10/ai-query-electricity-use-scenario-calculation-unhyd.png) A practical job-offer check separates three questions: does the role exist, is the recruiter authorized, and is the request for sensitive onboarding data appropriate? Failure at any stage means pausing and verifying independently. ## Why a text prompt is not a video, a research agent or a training run Short text answers, long-context reasoning, image generation, video generation and multi-step agents do different amounts of computational work. They can also use different models, accelerator hardware and serving arrangements. Treating all of these as one standard 'AI query' conceals precisely the variables that matter. A [Microsoft Research-led paper](https://arxiv.org/abs/2509.20241?ref=unhyd.com) developed a bottom-up model of inference energy. For its specified frontier-scale system, it estimated a median of 0.34 Wh per query, but a test-time scaling scenario involving much more generated computation raised its modeled median to 4.32 Wh. Those values are estimates under stated hardware and workload assumptions, not measurements of ChatGPT, Gemini or all AI agents. The useful finding is the sensitivity to task design, not a universal price tag for reasoning. The IEA's 2026 report makes the same distinction at the system level: simple text-query energy is only one component of a rapidly evolving mix. The agency warned that advanced reasoning, agents and video generation can be hundreds or thousands of times more energy-intensive per task than simple text generation in some cases. It also reported strong efficiency improvements. Both workload growth and efficiency change are necessary to understand the direction of total demand. ## The boundary problem: what gets counted and what is left out Any responsible estimate needs a clear boundary. An accelerator-only measurement excludes host processors and memory. A server-only measurement excludes cooling and electrical distribution overhead. A full data-center serving estimate can include idle capacity and facility overhead, but it still does not automatically include chip manufacturing, construction, training a new model, electricity generation's upstream impacts or every network and end-user device involved in a task. Google's 2025 paper explicitly included several production-serving components that narrower GPU benchmarks omit. That is a strength for its stated purpose, but it does not turn the result into a full lifecycle assessment. Google's 0.26-milliliter water figure for a median text prompt also uses a particular accounting scope; it should not be equated with all water consumed throughout the electricity supply chain. A water-use claim therefore needs its own definition of what is counted, rather than borrowing the electricity boundary without checking it. Even carbon emissions require a different calculation from electricity. A watt-hour measures energy, not emissions. Emissions depend on the electricity supply, accounting method and, for lifecycle analysis, the manufacture and operation of equipment. Two systems using the same electrical energy need not have identical carbon footprints. ## Why a falling cost per prompt can coexist with rising data-center demand The [IEA's April 2026 report](https://www.iea.org/reports/key-questions-on-energy-and-ai?ref=unhyd.com) says data-center electricity consumption worldwide rose 17% in 2025, while electricity use by AI-focused data centers rose 50%. These are reported sector-level changes, not per-query measurements. The agency also describes a surge in AI-related investment and rapidly changing demand for higher-density computing equipment. The apparent paradox disappears once the variables are separated. Total energy depends on how many tasks are run, what each task requires, how efficiently infrastructure handles them and which other activities a data center performs. A service can reduce electricity per typical request while increasing total electricity because it serves many more requests or shifts toward heavier work. That is not proof that efficiency improvements failed; it is a reminder that unit efficiency and aggregate consumption answer different questions. For a separate view of grid-scale uncertainty, Unhyd's [data-center electricity forecast analysis](https://unhyd.com/article/data-center-electricity-forecast-2030-explained/) explains why projections for 2030 are ranges rather than guaranteed outcomes. And the publication's [green software overview](https://unhyd.com/article/green-software-engineering-carbon-footprint/) describes why application design and infrastructure efficiency are connected. Neither page is a substitute for measuring the actual workload under discussion. ## Five questions to ask before sharing an AI energy claim - **Which product and task?** A median short text prompt is not a long autonomous research run. - **Which statistic?** Median, mean, benchmark estimate and measured fleet total cannot be swapped. - **Which components?** Check whether the boundary includes accelerator, host, idle capacity and cooling overhead. - **Which date and hardware?** Efficiency, model routing and demand change; a 2025 value is not a guaranteed 2026 value. - **What is being inferred?** A personal-use illustration is not evidence of national electricity demand or lifecycle emissions. Those questions offer a better reading of the numbers than a contest over whose chatbot supposedly uses fewer seconds of household electricity. The most defensible conclusion is narrower: published figures suggest some ordinary text prompts can be served with fractions of a watt-hour under particular conditions, while much larger and more complex AI infrastructure is consuming a growing amount of electricity. The gap between those claims is not a contradiction. It is the measurement problem that serious comparisons must solve.