How much energy does AI use? A single normal text prompt uses roughly 0.24 watt-hours, Google’s official figure for a median Gemini prompt, to about 0.34 watt-hours for an average ChatGPT query. That is about what an oven uses in one second. The model you pick changes everything. An independent 2025 benchmark put Claude Sonnet at roughly 2.8 Wh for a medium prompt, DeepSeek-V3 near 2 Wh for a medium prompt, and DeepSeek-R1 reasoning at about 29 Wh, the same heavyweight tier as GPT-5-class models. Credible per-prompt estimates span more than 65x. At scale the picture is enormous, with global data centers consuming about 415 terawatt-hours of electricity in 2024 and projected to roughly double by 2030.
This guide gives you verified numbers for every version of the question, so per prompt, per model, per year, and per data center, with Gemini, ChatGPT, Claude, DeepSeek, Grok and open models all measured side by side. Every figure is dated and attributed to its source. You will see why the cheapest model to buy is not the cheapest to run, why credible estimates differ so wildly, and which viral claims were inflated. You will also see how AI energy use compares to a Google search or an hour of streaming, and the single change that actually lowers your footprint. We lead with data, not panic, and we tell you what the data does not yet show.
The Key Takeaways
- A normal prompt is tiny: ~0.24 to 0.34 Wh. Google’s official 2025 figure for a median Gemini text prompt is 0.24 Wh; OpenAI’s Sam Altman gives 0.34 Wh for an average ChatGPT query, about what an oven uses in one second.
- Cheap to buy is not cheap to run. In one consistent 2025 benchmark, DeepSeek-R1 used about 29 Wh per prompt while DeepSeek-V3 stayed near 2 Wh, versus roughly 2.8 Wh for Claude Sonnet and 1.2 Wh for GPT-4o. DeepSeek’s low API price hides high inference energy.
- Reasoning models cost 10 to 65x more. A medium GPT-5-class response is estimated at ~18 Wh and up to 40 Wh, the same tier as DeepSeek-R1 y o3, which is why credible estimates span more than 65x across models.
- Grok and GPT-5.5 have no first-party number. Neither xAI nor OpenAI publishes a per-query energy figure for their current flagships, so any specific Grok or GPT-5.5 watt-hour claim is an estimate, not disclosure.
- Global data centers used about 415 TWh in 2024, roughly 1.5% of world electricity; the IEA projects this roughly doubles to ~945 TWh by 2030, close to the entire annual electricity use of Japan.
- Model choice is the biggest lever you control. Sending a simple question to a frontier reasoning model can use roughly 10x to 65x the energy of a lean model giving an equally good answer.
How much energy does AI use per prompt?
For a normal text question on a lean model, the honest answer is roughly 0.24 to 0.34 watt-hours. The low end is Google’s official 2025 figure for a median Gemini prompt; the high end is the number OpenAI CEO Sam Altman published for an average ChatGPT query. Both are tiny, about what a microwave draws in one second. The scarier numbers come from a different reality. The model and the prompt length change the answer by orders of magnitude, so a single “per prompt” figure hides a spread of more than 65x once Claude, DeepSeek and reasoning models are in the table.
A Comparison Table
| Model / estimate | Energy per prompt | Roughly equal to | What’s counted | Source & year |
|---|---|---|---|---|
| Google Gemini (median text prompt) | ~0.24 Wh | a laptop for ~17 seconds | Cooling + full data-center overhead | Google, Aug 2025 |
| ChatGPT (average query) | ~0.34 Wh | an oven for ~1 second | OpenAI’s stated figure | Sam Altman, Jun 2025 |
| GPT-4o (typical query) | ~0.3 Wh | one Google search | GPU compute estimate | Epoch AI, Feb 2025 |
| Claude Sonnet (medium prompt) | ~2.8 Wh (0.8 to 5.5) | a 60W bulb for ~3 minutes | Independent inference benchmark | Jegham et al., May 2025 |
| DeepSeek-V3 (medium prompt) | ~2 Wh | a 60W bulb for ~2 minutes | Independent inference benchmark | Jegham et al., May 2025 |
| DeepSeek-R1 (reasoning) | ~29 Wh | ~120 lean Gemini prompts | Independent inference benchmark | Jegham et al., May 2025 |
| GPT-5 / reasoning class (medium) | ~18 Wh (up to 40) | a lightbulb for ~2 hours | GPU compute estimate | URI AI Lab, Aug 2025 |
| Grok / GPT-5.5 (current flagships) | Not disclosed | n/a | No first-party figure published | xAI / OpenAI, 2026 |
One caveat keeps this honest. The first-party Google and OpenAI numbers measure a lean median prompt; the independent “How Hungry is AI?” benchmark by Jegham and colleagues measures realistic medium and long prompts across many models with full overhead, which is why its figures sit higher. They are not contradictory, they are different scopes, and the section below shows how to read both.
Google Gemini: about 0.24 Wh per prompt
In August 2025 Google published the most detailed first-party measurement so far, based on production data from May 2025. It found that a median Google Gemini text prompt uses 0.24 watt-hours of energy and emits 0.03 grams of CO2 equivalent. Crucially, that figure is not a cherry-picked best case, because Google’s method includes cooling, idle reserve capacity, CPU and RAM, and data-center overhead, which makes it one of the more honest single numbers available. You can read the method in Google’s environmental impact report.
ChatGPT and GPT-5.5: about 0.34 Wh per query, flagship undisclosed
In a June 2025 blog post, OpenAI CEO Sam Altman wrote that an average ChatGPT query uses 0.34 watt-hours, what an oven would use in a little over one second. It is OpenAI’s first official datapoint, it has not been peer reviewed, and the company never defined what counts as an “average” query. Independent work supports the order of magnitude, with research institute Epoch AI estimating a typical GPT-4o query at roughly 0.3 Wh.
That 0.34 Wh number predates the current flagship. GPT-5.5 shipped in April 2026 and became the default ChatGPT model in May 2026, yet OpenAI has published no per-query energy figure for it. Since GPT-5.5 routes more queries through reasoning, the realistic expectation is higher than 0.34 Wh for anything that triggers extended thinking, closer to the heavy tier below than to the lean median.
Claude: no official figure, roughly 1 to 5 Wh in benchmarks
Anthropic does not publish an energy-per-prompt figure for Claude, so every Claude number is an external estimate. The cleanest is the Jegham et al. inference benchmark (May 2025), which measured Claude Sonnet at about 0.8 Wh for a short prompt, 2.8 Wh for a medium one, and 5.5 Wh for a long one. Stanford’s 2026 AI Index reaches the same conclusion from a different method, placing a Claude flagship near the efficient end of frontier models for a medium prompt. The same benchmark rated Claude Sonnet the most eco-efficient model overall once answer quality is held constant. Claude delivers strong answers without the runaway energy bill of a pure reasoning model.
DeepSeek: cheap to buy, expensive to run
DeepSeek is the clearest example of why price and energy are not the same thing. Its API is among the cheapest of any frontier model. Yet the Jegham benchmark measured DeepSeek-V3 at about 2 Wh for a medium prompt, while the reasoning-tuned DeepSeek-R1 hit 29 Wh, among the most energy-intensive models tested and well above Claude or GPT-4o. Stanford’s 2026 AI Index found the same pattern, with a DeepSeek model using several times the per-prompt energy of a Claude flagship for comparable work.
The newer DeepSeek V4, released in late April 2026, uses a 1.6-trillion-parameter mixture-of-experts design with only about 49 billion parameters active per token (and a 284B Flash variant with ~13B active), which DeepSeek says cuts inference cost sharply. That is a credible efficiency story on paper, but no independent per-query watt-hour figure for V4 exists yet, so treat any specific V4 energy claim as unverified for now.
Grok: undisclosed per prompt, data-center scale is the real signal
xAI publishes no per-query energy figure for Grok, and credible third-party ranges are wide, from a fraction of a watt-hour for a simple query to several watt-hours for heavy reasoning or agentic runs. The more concrete Grok signal is infrastructure. xAI’s Colossus supercomputer in Memphis has been reported in the hundreds of megawatts, and satellite analyses put the newer Colossus 2 in the gigawatt range, the scale at which a single AI cluster rivals a mid-sized city’s electricity draw. For Grok, the honest answer is that the per-prompt number is unknown and the systems-level footprint is large and growing fast.
Open models: the low end of the range
Smaller open models anchor the efficient end. In the same benchmark, a 1-billion-parameter Llama 3.2 used about 0.07 Wh for a short prompt, while the very large Llama 3.1 405B used about 9 Wh for a medium one. The takeaway is consistent across providers, so size and reasoning depth drive energy far more than which company built the model, which is exactly why matching the model to the task matters.
Why do AI energy estimates disagree so much?
Most of the confusion comes from what each number counts. A low figure like Google’s 0.24 Wh measures the electricity to run the chips plus cooling and overhead for one median lean prompt. A scary figure counts a heavyweight reasoning model, a long prompt, the full data-center power draw, and sometimes the energy to build the hardware. Different scopes produce numbers that differ by more than 65x while every one of them is technically correct.
Model architecture is the second driver. A dense model runs every parameter for every token, while a mixture-of-experts model like DeepSeek V4 activates only a slice, so two models with similar total size can differ wildly in energy. Reasoning modes add a third layer. A model that “thinks” before answering can generate several times more tokens than a direct reply, which is why DeepSeek-R1 and o3-class models land near 20 to 40 Wh while GPT-4o stays near 1 Wh on the same test.
The last driver is training versus inference. Training is a one-time, very large run; training GPT-3 used an estimated 1,287 megawatt-hours, and training GPT-4 is estimated near 50 gigawatt-hours. Inference is the electricity spent every time you send a prompt, and because a popular model answers billions of prompts a day, inference now dominates total AI energy use. The widely quoted 2.9 Wh “ten times a Google search” figure dates to early estimates that were later revised down by roughly an order of magnitude, which is why it keeps colliding with newer first-party numbers. This is one of the most persistent common myths about AI energy use.
How much energy do AI data centers use?
Per prompt the numbers are tiny, but the aggregate is national-scale. According to the IEA’s Energy and AI report, global data centers consumed about 415 terawatt-hours of electricity in 2024, roughly 1.5% of world electricity. The IEA’s base case projects this roughly doubles to about 945 TWh by 2030, just under 3% of global electricity and close to Japan’s entire annual use. AI-accelerated servers are the fastest-growing piece, at around 30% a year.
| Metric | Figure | Year | [fuente] |
|---|---|---|---|
| Global data center electricity | ~415 TWh (~1.5% of world total) | 2024 | IEA |
| Global data centers, projected | ~945 TWh (~3% of world total) | 2030 | IEA, base case |
| ChatGPT, estimated annual energy | ~310 GWh (≈ 29,000 US homes) | 2025 | IEEE Spectrum |
| All generative AI | ~15 TWh | 2025 | IEEE Spectrum |
| All generative AI, projected | ~347 TWh | 2030 | IEEE Spectrum |
| US data centers, share of US electricity | ~4.4% | 2023 | MIT Technology Review |
You will also see headlines claiming data centers will hit more than 1,000 TWh by 2026. That figure comes from the IEA’s earlier 2024 electricity report, a faster near-term projection; the IEA’s more detailed 2025 Energy and AI report settled on the ~945 TWh by 2030 base case used above. The trajectory, not any single year’s headline, is the real story. In the US specifically, data centers already drew about 4.4% of all electricity in 2023, with more than half of that projected to go to AI by 2028.
Is that actually a lot? AI energy use compared to other things
For an individual, a lean AI prompt is a rounding error in your daily energy use. Google’s own framing puts a 0.24 Wh prompt at about the energy of watching TV for under nine seconds. The honest individual takeaway is that everyday lean chatbot use is negligible next to a single hot shower or a short drive. An hour of streaming video uses tens of watt-hours, easily outweighing hundreds of lean prompts. The exception is heavy reasoning, where a single DeepSeek-R1 or GPT-5-class response can equal more than a hundred lean prompts on its own.
| Activity | Energy | Equivalent |
|---|---|---|
| One lean AI prompt (Gemini) | ~0.24 Wh | a laptop for ~17 seconds |
| One Google search | ~0.3 Wh | similar to a lean AI prompt |
| One average ChatGPT query | ~0.34 Wh | an oven for ~1 second |
| One Claude Sonnet medium prompt | ~2.8 Wh | ~12 lean prompts |
| One DeepSeek-R1 reasoning prompt | ~29 Wh | ~120 lean prompts |
| One heavy GPT-5-class response | ~18 Wh | ~75 lean prompts |
| Streaming 1 hour of video | tens of Wh | hundreds of lean prompts |
The catch is scale and concentration. Individually negligible prompts add up to ~15 TWh across all generative AI in 2025, projected near 347 TWh by 2030, and that load lands on specific regional grids rather than spread evenly. So the personal-footprint answer (“barely anything”) and the systems answer (“a fast-growing strain on the grid”) are both true, which is exactly why the topic is so easy to spin in either direction.
Is AI getting more energy efficient?
Per prompt, yes, and fast. Google reported that the energy of a median Gemini prompt fell roughly 33-fold in twelve months through better models, hardware, and software. Mixture-of-experts designs like DeepSeek V4 push the same direction by activating only a fraction of parameters per token. The problem is the Jevons effect, where cheaper, better AI gets used far more, so total consumption keeps climbing even as each prompt gets cleaner.
That tension is the core of where AI is heading in 2026. Efficiency gains are real and large, but agentic systems change the math, because one request from an AI agent can fan out into dozens of chained model calls. Lean per-prompt numbers do not capture a workflow that quietly runs a reasoning model fifty times to complete one task, which is the next measurement gap the public data has barely started to address.
The one lever you control: model choice
The single most effective thing an individual controls is which model answers the question, because the spread between models is enormous. Sending a simple query to a frontier reasoning model can use roughly 10x to 65x the energy of a lean model giving an equally good answer, and as the table above shows, that gap holds across Claude, DeepSeek, Gemini and GPT-class models. Matching the model to the task does more than any amount of prompt-trimming. Comparisons like DeepSeek vs ChatGPT, ChatGPT vs Gemini, and our guide to the most capable AI models show how different the efficiency profiles really are.
Practically, that means defaulting to a lighter model for everyday questions and reserving heavyweight reasoning models for problems that need them. Running several models from one place makes “pick the right tool” realistic instead of theoretical. That is the point of a multi-model setup like the one in our getting started guide, where one subscription gives you Claude, ChatGPT, Gemini, Grok, and DeepSeek side by side. The lean option is one tap away from the heavy one. None of this requires guilt; it just means using the heavy option deliberately rather than by default.
Conclusión
So how much energy does AI use? A normal prompt on a lean model is realistically 0.24 to 0.34 watt-hours, but the model matters more than anything, with Claude Sonnet near 2.8 Wh, DeepSeek-R1 near 29 Wh, and GPT-5-class reasoning at 18 Wh or more, a spread of over 65x. At scale the numbers are large and growing fast, with global data centers at about 415 TWh in 2024 and projected to roughly double by 2030. The most useful next step is not to stop using AI; it is to be deliberate about which model you reach for, starting with a clear-eyed look at how the leading models actually differ.
FAQ
How much energy does one ChatGPT prompt use?
About 0.34 watt-hours for an average query, the figure OpenAI CEO Sam Altman published in June 2025, roughly what an oven uses in one second. That number predates GPT-5.5, which is now the default and has no published per-query figure, so reasoning-heavy replies likely use considerably more.
Which AI model uses the most energy?
Among widely tested models, reasoning models lead. In one consistent 2025 benchmark, DeepSeek-R1 used about 29 Wh and o3-class models reached 20 to 40 Wh for medium to long prompts, versus roughly 2.8 Wh for Claude Sonnet and about 1 Wh for GPT-4o. DeepSeek is cheap to buy but expensive to run.
Does an AI prompt use more energy than a Google search?
Not for lean models. First-party 2025 figures put a normal AI prompt at roughly 0.24 to 0.34 Wh, about the same as a single web search. The old “ten times a Google search” claim came from early estimates that were later revised down. Heavy reasoning prompts, however, can use far more than a search.
How much electricity do AI data centers use?
Global data centers used about 415 TWh in 2024, roughly 1.5% of world electricity. The IEA projects this roughly doubles to about 945 TWh by 2030, close to Japan’s entire national use, with AI the fastest-growing driver.
Which AI model is most energy efficient?
Lean, smaller models are far more efficient than frontier reasoning models, often by 10x or more for a comparable everyday answer. Among major chat models, Claude Sonnet scored best for energy efficiency once answer quality is held constant. The biggest saving you control is matching the model to the task.
