Explainer
How much energy an AI query uses
Company and outside estimates put a typical text prompt to a chatbot at roughly 0.24 to 0.34 watt-hours of electricity, similar to the 0.3 Wh Google gave for a web search in 2009.[1][2][3][4] Figures vary with what is counted and how long the task is: Google's estimate falls to 0.10 Wh if only the AI chips are counted, and long or reasoning-heavy queries can use far more.[5][6] Per query the energy is small, but the IEA estimates data centres used about 485 TWh in 2025 and could reach about 950 TWh by 2030.[7][8]
Most per-prompt figures on this page are self-reported by AI companies or estimated by outside researchers, and none is independently audited. Compare them with care. For the big picture on data centre electricity, read AI data centers and electricity, explained. Choose a reading level in each section.
The headline numbers
Google estimated in August 2025 that a typical text question to its Gemini app uses 0.24 Wh, which it compared to watching TV for less than nine seconds.[1]
OpenAI’s chief executive wrote in June 2025 that an average ChatGPT question uses about 0.34 Wh, in a blog post.[2] Epoch AI, a research group, estimated roughly 0.3 Wh for a typical ChatGPT question using GPT-4o.[3]
For comparison, Google said in 2009 that a web search used about 0.3 Wh.[4]
| Source | Figure | Scope |
|---|---|---|
| Google (Aug 2025) | 0.24 Wh, 0.03 g CO2e, 0.26 mL water | Median Gemini Apps text prompt, full serving stack[1] |
| OpenAI CEO (Jun 2025) | ~0.34 Wh | Average ChatGPT query, stated in a blog post[2] |
| Epoch AI (Feb 2025) | ~0.3 Wh | Typical GPT-4o query, outside estimate[3] |
| Google (2009) | 0.3 Wh | Average web search, including index building[4] |
Google published the methodology behind its figure in an August 2025 paper.[5]
Why the estimates differ
It depends on what you count. Google’s 0.24 Wh includes the AI chips, the ordinary computers around them, spare machines kept running for busy moments, and the cooling and power systems of the data centre. If you count only the AI chips, the figure drops to 0.10 Wh.[5]
It also depends on the question. Epoch AI warns that very long questions, long answers, or “reasoning” models that think step by step can use much more than 0.3 Wh.[6]
Google’s paper accounts for active accelerator power, host system energy, idle machine capacity and data-centre overhead. Restricting the boundary to active TPU and GPU power yields 0.10 Wh, 0.02 g CO2e and 0.12 mL, which Google calls an optimistic underestimate.[5] Epoch AI cautions that long inputs, long outputs and reasoning models can push per-query energy well above 0.3 Wh.[6] It judges the older, widely repeated estimate of about 3 Wh per ChatGPT query, ten times a Google search, to be ten times too high.[3]
Water and carbon
Google estimated 0.26 mL of water, about five drops, for a typical Gemini text prompt.[1]
Mistral AI, a French company, published a broader study that also counts the impact of making the servers. It estimated that a 400-token answer from its Le Chat assistant produces 1.14 grams of CO2e and uses 45 mL of water.[9][10]
Google reports 0.03 g CO2e and 0.26 mL of water per median prompt.[1] Mistral’s lifecycle assessment with Carbone 4 and ADEME is broader, including upstream manufacturing, and reports 1.14 g CO2e and 45 mL of water per 400-token response, plus 20.4 kt CO2e for Mistral Large 2 as of January 2025, after 18 months of use.[9][10] The two studies draw their boundaries differently. See Mistral AI.
Model types and efficiency gains
Not all AI tasks cost the same. A 2023 study found that general-purpose generative AI models use far more energy than smaller models built for one task.[11]
AI is also getting more efficient. Google said the energy used by a typical Gemini text prompt fell 33 times over 12 months.[12]
Researchers at Hugging Face and Carnegie Mellon measured energy per 1,000 inferences and found multi-purpose generative architectures orders of magnitude more expensive than task-specific systems across a range of tasks.[11] Google reports a 33-fold fall in energy and a 44-fold fall in carbon per median prompt over a recent 12-month period.[12]
From one query to the whole grid
The IEA estimates that even if every ordinary web search became a simple AI question, the total would be less than 1% of data centre electricity.[7] Data centres overall are still growing fast: the IEA estimates they used about 485 TWh in 2025 and could use about 950 TWh in 2030.[8]
The IEA’s scenario of all conventional searches replaced by simple AI text queries gives under 4 TWh a year, below 1% of current data-centre consumption.[7] The IEA puts data centres at about 415 TWh in 2024 and 485 TWh in 2025, with about 950 TWh projected for 2030.[13][8] Those trends are covered in AI data centers and electricity, explained.
Read next
AI data centers and electricity, explained
How much electricity AI data centers use, why AI chips need so much power, where it comes from, cooling and water, and why grids struggle to keep up.
Developing in Science & Energy
Questions readers ask
How much electricity does one ChatGPT or Gemini prompt use?
Google estimated in 2025 that the median Gemini Apps text prompt uses 0.24 Wh. OpenAI's chief executive put the average ChatGPT query at about 0.34 Wh, and Epoch AI estimated roughly 0.3 Wh for a typical GPT-4o query. These are company or independent estimates, not audited figures.[1][2][3]
Does an AI query use ten times more energy than a Google search?
That comparison comes from an older estimate of about 3 Wh per ChatGPT query. Epoch AI's 2025 estimate of roughly 0.3 Wh is ten times lower, about the same as the 0.3 Wh Google reported per search in 2009.[3][4]
How much water does an AI prompt use?
Google estimated 0.26 mL of water for a median Gemini text prompt, about five drops. Mistral's lifecycle study estimated 45 mL for a 400-token Le Chat response, a figure that includes upstream impacts such as manufacturing servers.[1][9][10]
Why do estimates differ so much?
They count different things. Google's 0.24 Wh includes idle capacity, host CPUs and data-centre overhead; counting only active AI chips gives 0.10 Wh. Prompt length, output length and reasoning also change the energy used.[5][6]
Is AI energy use getting more efficient?
Google said the energy per median Gemini text prompt fell 33-fold over a recent 12-month period. Total data centre electricity use is still projected to rise.[12][8]
Do general-purpose AI models use more energy than task-specific ones?
A 2023 study by Hugging Face and Carnegie Mellon researchers found that multi-purpose generative models are orders of magnitude more expensive to run than task-specific models for a variety of tasks.[11]
Sources
· 13 claims from 8 publishers
Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.
- [1]
Google estimated in August 2025 that the median Gemini Apps text prompt uses 0.24 watt-hours of energy, emits 0.03 g CO2e and consumes 0.26 mL of water, which it compared to watching TV for less than nine seconds (company figure). confirmedas of 2025-08-21
- Measuring the environmental impact of AI inference · Google Cloud Blog · Primary source · 2025-08-21 · Introduction (retrieved 2026-11-07)
- Measuring the environmental impact of AI inference · Google Cloud Blog · Primary source · 2025-08-21 · Introduction (retrieved 2026-11-07)
- Measuring the environmental impact of AI inference · Google Cloud Blog · Primary source · 2025-08-21 · Introduction (retrieved 2026-11-07)
- Measuring the environmental impact of AI inference · Google Cloud Blog · Primary source · 2025-08-21 · Introduction (retrieved 2026-11-07)
- Measuring the environmental impact of delivering AI at Google Scale · arXiv (Google authors) · Primary source · 2025-08-21 · Abstract (retrieved 2026-11-07)
- [2]
OpenAI's chief executive wrote in June 2025 that the average ChatGPT query uses about 0.34 watt-hours, "about what an oven would use in a little over one second", in a parenthetical note on his blog. reportedas of 2025-06-10
- The Gentle Singularity · Sam Altman (blog) · Secondary source · 2025-06-10 · Blog post, parenthetical note (retrieved 2026-11-07)
- [3]
Epoch AI estimated in February 2025 that a typical ChatGPT query using GPT-4o uses roughly 0.3 watt-hours, ten times less than the commonly cited 3 watt-hour figure, which was also said to be ten times a Google search. reportedas of 2025-02-07
- How much energy does ChatGPT use? · Epoch AI · Secondary source · 2025-02-07 · Introduction (retrieved 2026-11-07)
- How much energy does ChatGPT use? · Epoch AI · Secondary source · 2025-02-07 · Introduction (retrieved 2026-11-07)
- [4]
In 2009 Google said an average search used 0.0003 kWh (0.3 Wh, about 1 kJ) of energy, including work such as building the search index. confirmedas of 2009-01-11
- Powering a Google search · Official Google Blog · Primary source · 2009-01-11 · Body (retrieved 2026-11-07)
- [5]
Google's figure counts active accelerator power, host CPU and memory, idle capacity kept ready for traffic spikes and data-centre overhead; counting only active TPU and GPU power gives 0.10 Wh, 0.02 g CO2e and 0.12 mL of water, which Google calls an optimistic underestimate. confirmedas of 2025-08-21
- Measuring the environmental impact of delivering AI at Google Scale · arXiv (Google authors) · Primary source · 2025-08-21 · Abstract (retrieved 2026-11-07)
- Measuring the environmental impact of AI inference · Google Cloud Blog · Primary source · 2025-08-21 · Calculating the footprint (retrieved 2026-11-07)
- Measuring the environmental impact of AI inference · Google Cloud Blog · Primary source · 2025-08-21 · Calculating the footprint (retrieved 2026-11-07)
- Measuring the environmental impact of AI inference · Google Cloud Blog · Primary source · 2025-08-21 · Idle machines (retrieved 2026-11-07)
- [6]
Epoch AI cautioned that queries with long inputs or long outputs, such as those using reasoning models, may use substantially more than 0.3 watt-hours. reportedas of 2025-02-07
- How much energy does ChatGPT use? · Epoch AI · Secondary source · 2025-02-07 · Conclusion (retrieved 2026-11-07)
- [7]
The IEA estimates that if all conventional internet searches were done as simple AI text queries, they would use less than 4 TWh a year, under 1% of current data centre consumption. confirmedas of 2026-04-16
- Key Questions on Energy and AI - Executive summary · International Energy Agency · Primary source · 2026-04-16 · Executive summary (retrieved 2026-11-07)
- [8]
The IEA's April 2026 projection has data centre electricity use roughly doubling from 485 TWh in 2025 to 950 TWh in 2030, about 3% of global electricity demand. confirmedas of 2026-04-16· forecast
- Key Questions on Energy and AI - Executive summary · International Energy Agency · Primary source · 2026-04-16 · Section: Our updated data centre electricity demand outlook (retrieved 2026-11-07)
- [9]
In July 2025 Mistral published a lifecycle analysis, done with Carbone 4 and the French ecological transition agency ADEME, estimating that, as of January 2025 and after 18 months of use, Mistral Large 2 had produced 20.4 kilotonnes of CO2e and consumed 281,000 cubic metres of water, and that a 400-token Le Chat response produces 1.14 g CO2e and uses 45 mL of water. confirmedas of 2025-07-22
- Our contribution to a global environmental standard for AI · Mistral AI · Primary source · 2025-07-22 · The first comprehensive lifecycle analysis (retrieved 2026-11-07)
- Our contribution to a global environmental standard for AI · Mistral AI · Primary source · 2025-07-22 · Disclosed metrics (retrieved 2026-11-07)
- Our contribution to a global environmental standard for AI · Mistral AI · Primary source · 2025-07-22 · Disclosed metrics (retrieved 2026-11-07)
- [10]
Mistral says its lifecycle figures include upstream emissions such as manufacturing servers, not just energy use. confirmedas of 2025-07-22
- Our contribution to a global environmental standard for AI · Mistral AI · Primary source · 2025-07-22 · Disclosed metrics (retrieved 2026-11-07)
- [11]
A 2023 study by researchers at Hugging Face and Carnegie Mellon, measuring the energy of 1,000 inferences, found multi-purpose generative models are orders of magnitude more expensive to run than task-specific models for a variety of tasks. confirmedas of 2024-10-15
- Power Hungry Processing: Watts Driving the Cost of AI Deployment? · arXiv (Hugging Face and Carnegie Mellon authors) · Primary source · 2023-11-28 · Abstract (retrieved 2026-11-07)
- [12]
Google said the energy and total carbon footprint of the median Gemini Apps text prompt fell 33-fold and 44-fold over a recent 12-month period. confirmedas of 2025-08-21
- Measuring the environmental impact of AI inference · Google Cloud Blog · Primary source · 2025-08-21 · Introduction (retrieved 2026-11-07)
- [13]
The IEA estimates that data centres used around 1.5% of the world's electricity in 2024, or 415 TWh. confirmedas of 2025-04-10
- Energy and AI - Executive summary · International Energy Agency · Primary source · 2025-04-10 · Executive summary (retrieved 2026-11-07)
Revision history (2)
- Attributed the summary's per-prompt range to company and outside estimates.
- Page created.
Created Nov 7, 2026. Last reviewed by an editor on Nov 8, 2026. Next scheduled review: Feb 7, 2027.
Cite this page
"How much energy an AI query uses." ContentLora, updated Nov 14, 2026. https://contentlora.com/explain/ai-energy-per-query-explained
Spotted an error? Suggest a correction or emailcorrections@contentlora.com.