Skip to content

How much does a Copilot prompt consume?

A Copilot prompt most likely uses about 0.31 Wh of electricity, somewhere in a 0.16 to 0.60 Wh band. The figure comes from Microsoft, and it does not measure Copilot: it is the median of a simulation of large-scale inference, published in Joule in April 2026 (checked November 11, 2026). In the same simulation a reasoning query reaches 3.91 Wh, 13 times more.

0.31 Wh
Simulated median, interquartile range 0.16 to 0.60 Wh: no vendor figure for this product

Research verified on November 11, 2026 · Sources: Oviedo et al. : « Energy use of AI inference, efficiency pathways, and test-time scaling », Joule (avril 2026, DOI 10.1016/j.joule.2026.102430 ; préprint arXiv 2509.20241) : médiane de 0,31 Wh par requête, intervalle interquartile 0,16 à 0,60 Wh · Microsoft Cloud Blog (15 juin 2026) : « Scaling AI with 8 to 20x energy efficiency », 0,16 à 0,60 Wh et 0,0 à 0,067 mL d'eau pour une requête type · Google Cloud (21 août 2025) : impact environnemental de l'inférence IA, 0,24 Wh, 0,26 mL d'eau et 0,03 g CO₂e pour le prompt texte médian de Gemini Apps · OpenAI / Sam Altman : « The Gentle Singularity » · Epoch AI : énergie par requête ChatGPT · « How Hungry is AI? » (arXiv 2505.09598) · de Vries : « The growing energy footprint of artificial intelligence », Joule (oct. 2023, DOI 10.1016/j.joule.2023.09.004) : 2,9 Wh par requête ChatGPT contre 0,3 Wh pour une recherche Google, hypothèse fondée sur un parc de GPU Nvidia A100

Energy of one text query: the figures Microsoft’s study compares with its own median
SourceWhat the figure isPer queryBetween 0.16 and 0.60 Wh?
Microsoft (Oviedo et al., Joule, April 2026)Simulation on three open-weight models above 200 billion parameters, median of 300 output tokens0.31 WhThe study’s median
Google, Gemini Apps (August 2025)Measured on the vendor’s own fleet0.24 WhYes
OpenAI, ChatGPT (June 2025)Stated figure, no published method0.34 WhYes
Epoch AI, GPT-4o (2025)Independent estimate from compute, 500 output tokens0.30 WhYes
“How Hungry is AI?”, GPT-4o (2025)Independent estimate from observed throughput, 300 output tokens, Microsoft Azure servers0.42 WhYes
de Vries, ChatGPT (Joule, 2023)Estimate under non-production conditions2.90 WhNo

Source: Table 1 of Oviedo et al., preprint arXiv 2509.20241 (version 2, 9 June 2026), read on November 11, 2026. Values are those the paper reports, for a text query without reasoning. The authors note that the comparison is not standardised: products, models, query lengths and methods differ. None of these rows measures Copilot.

Where does the 0.31 Wh come from?

From a peer-reviewed paper by eight Microsoft researchers: Felipe Oviedo and co-authors, “Energy use of AI inference, efficiency pathways, and test-time scaling”, Joule, April 2026 (DOI 10.1016/j.joule.2026.102430; preprint arXiv 2509.20241, posted in September 2025 and re-read for this page in its second version, dated 9 June 2026). No power meter was clamped to a server. The method is a Monte Carlo simulation with 10,000 draws per model. The authors start from published throughput for five open-weight models, in tokens per second on nodes of eight or ten H100 GPUs. Their typical query has 500 input tokens and a median of 300 output tokens. The 0.31 Wh figure pools the three models above 200 billion parameters: DeepSeek-R1, Llama 3.1 405B and Llama 3.1 Nemotron Ultra. The output is a distribution, not a single number: median 0.31 Wh, interquartile range 0.16 to 0.60 Wh. The code released alongside the paper calls these results “directional” rather than definitive measurements, and says what they leave out: no hardware manufacturing, no full life-cycle accounting.

Why does that figure not measure Copilot?

Because Microsoft never claims it does. The word Copilot does not appear once in the paper. The Microsoft Cloud blog post of 15 June 2026 names it exactly once, in its opening line, as an example of a service built on a large language model. That is an illustration, not an attribution. Copilot does not run on the open-weight models the study simulates. It runs on OpenAI models and on Microsoft’s own models, whose real throughput is not public. The only bridge the authors build is an assumption about size: they consider the 200-billion-parameter threshold a reasonable range for closed models such as GPT-4o. The simulation brackets Copilot; it does not weigh it. The situation mirrors Claude, where no vendor figure exists either, except that the bracket for Copilot is far better documented.

Does it agree with Google's and OpenAI's numbers?

Yes, and that is what makes it useful. The authors publish the comparison themselves (Table 1 of the paper), and the table on this page reproduces it. Google measured 0.24 Wh for the median text prompt in Gemini Apps on 21 August 2025, on its own fleet. OpenAI stated 0.34 Wh for an average query on 10 June 2025, with no method note. Two independent estimates for GPT-4o give 0.30 Wh (Epoch AI) and 0.42 Wh (the “How Hungry is AI?” study, which places OpenAI’s models on Microsoft Azure servers). All four sit inside the 0.16 to 0.60 Wh range. The figure most quoted before them, 2.9 Wh per ChatGPT query (de Vries, 2023), does not: it is about nine times Microsoft’s median, a ratio computed by this site. The 0.3 Wh this site uses for a text prompt sits within 3% of that median. The authors argue that estimates made outside production overstate energy use by 4 to 20 times, because they leave out batching and inference acceleration. The argument is technical and it holds. It also comes from a company with an interest in a low number: all eight authors are Microsoft employees, and the paper says so.

What about water?

The same post gives 0.0 to 0.067 mL per typical query, with a median Microsoft describes as less than a single drop. The gap with Google is stark: 0.26 mL for Gemini, close to four times the top of Microsoft's range. Both figures cover water consumed on site to cool the servers, and neither counts the water withdrawn upstream to generate the electricity. The gap comes from the fleet and the climate: how much cooling is done with air, how efficiently water is used, and where the data centers sit weigh more than the model does. Comparing two water figures across vendors compares two geographies, not two AIs.

What changes with a reasoning query?

Everything, and this is the paper’s most quotable result. A reasoning query, which the authors set at a median of 5,000 output tokens, costs 3.91 Wh (interquartile range 2.15 to 7.05 Wh): 13 times the typical query. A sizeable share of the draws goes above 10 Wh. The authors then scale that up to a whole service. One billion typical queries a day need 0.73 GWh. If 10% of those queries switch to reasoning, the total climbs to 1.7 GWh a day, more than double. Copilot is pushing exactly those agent and reasoning modes. The 0.31 Wh median therefore describes yesterday’s usage, not where the product is heading. The same authors expect the opposite pull as well: efficiency gains in models, serving stacks and chips that could cut energy per query by 8 to 20 times. With cautious gains of 1.5 to 3 times, their 10% reasoning scenario falls back to 0.8 GWh a day. They warn themselves that an efficiency gain does not guarantee lower total consumption: a cheaper query invites more of them.

What does the simulation leave out when Copilot works as an agent?

Almost everything that separates an agent from a single question. Microsoft’s simulation defines a query as one prompt and one text reply. An agent that chains ten steps therefore counts as ten queries, and its cost follows the total number of tokens it generates. The authors list what they did not count: orchestration between steps, tool calls and the handling of long contexts. They also say their model handles very long inputs poorly, around 100,000 tokens, where reading the context takes far more energy. They name the uses this affects: coding agents, multi-document summaries and retrieval. For very long outputs they hold throughput constant although real systems slow down, which probably understates energy. And the study covers text only, not images or video. The 0.31 Wh therefore applies to a short question put to an assistant, not to a task handed to an agent.

A Copilot prompt in everyday actions

No official figure exists for Copilot: these equivalents use the site's standard assumption for a text prompt (0.3 Wh, 0.3 mL, 0.2 g), not vendor data.

This figure says what a prompt costs, not what it returns. For a small business the real question comes next: which Copilot habits earn the time they take, and which are pure reflex. That is the ground the ghis.fr studio works on, and it publishes this calculator. Here we measure and cite, and stop there.

The other models