How much electricity did your last AI question use?
You might expect there to be a simple answer by now. There isn’t.
One widely reported figure puts an average ChatGPT query at about 0.34 watt-hours. Google has measured the median text prompt in its Gemini app at 0.24 watt-hours. A 2026 study of frontier-scale AI systems estimated a typical query at a median of 0.31 watt-hours.
But other AI requests can use many times more energy. A long reasoning query may require several watt-hours, while measurements of individual models and workloads have produced results exceeding 20 watt-hours per prompt.
Those figures are not necessarily contradictions.
An AI prompt is not a standard unit of work. Asking a small model to answer “What is the capital of France?” is fundamentally different from asking a frontier reasoning model to analyse a legal document, write thousands of words, run tools and revise its answer repeatedly.
The hardware, model, output length, server utilisation, data centre and even the way researchers define “energy use” can all change the result.
This guide looks at what the best available research actually tells us about the energy used by artificial intelligence — and why simple claims such as “one AI prompt uses X amount of electricity” should be treated cautiously.
AI Energy Use at a Glance
| AI activity | Reported or estimated energy use | Important context |
|---|---|---|
| Median Gemini text prompt | 0.24 Wh | Google production measurement, May 2025 |
| Average ChatGPT query | About 0.34 Wh | Figure disclosed by OpenAI CEO Sam Altman |
| Typical frontier-scale text query | Median 0.31 Wh | 2026 production-scale modelling |
| Long reasoning query | Median 3.91 Wh | Approximately 5,000 output tokens in the study scenario |
| Individual large-model benchmarks | Can exceed 20 Wh | Depends heavily on model, prompt and benchmarking methodology |
| AI image generation | Highly variable | A 2025 study found up to a 46-fold difference between models |
The most important number in this table may not be any individual watt-hour figure. It is the amount of variation.
If you are new to this subject, our broader guide to AI sustainability explains the environmental impacts of artificial intelligence beyond electricity alone, including water use, hardware and emissions.
What Is a Watt-Hour?
Before comparing AI systems, it helps to understand the unit being used.
A watt measures power: how quickly energy is being used.
A watt-hour, written as Wh, measures an amount of energy. A device drawing one watt continuously for one hour consumes one watt-hour.
One kilowatt-hour, the unit commonly shown on household electricity bills, equals 1,000 watt-hours.
So when researchers say an AI request consumes 0.3 Wh, they are talking about a relatively small amount of electricity for that individual request.
The scale problem appears when requests are repeated millions or billions of times.
How Much Energy Does a ChatGPT Query Use?
In June 2025, OpenAI CEO Sam Altman wrote that the average ChatGPT query used about 0.34 Wh of electricity.
He compared this with the electricity used by a high-efficiency light bulb for a couple of minutes.
The number is useful because it came directly from OpenAI, but it has an important limitation: OpenAI did not publish a detailed measurement methodology alongside the figure.
We therefore do not know precisely what was considered an “average” query, which models were included, how long the average response was or exactly which infrastructure overheads were counted.
That does not make the number wrong. It does mean it should not be treated as a universal physical constant for ChatGPT.
AI systems also change rapidly. Routing systems may send simple questions to smaller models and more difficult requests to more capable models. Hardware improves, software becomes more efficient and user behaviour changes.
The energy required for a typical AI request in 2026 may therefore differ significantly from the energy required for an apparently similar request a year earlier.
Google Has Published a More Detailed Measurement
Google provided one of the clearest public examples of production AI energy measurement in 2025.
Its researchers reported that the median Gemini Apps text prompt consumed 0.24 Wh in a May 2025 measurement.
The same prompt was estimated to produce 0.03 grams of carbon dioxide equivalent and consume 0.26 millilitres of water.
Importantly, Google did not measure only the electricity flowing through the AI accelerator while it generated the answer.
Its methodology also included:
- energy used by active AI processors;
- host CPUs and memory;
- idle server capacity kept available for reliability and traffic spikes;
- data-centre cooling and power infrastructure; and
- the utilisation actually achieved in a real production environment.
Google calculated that looking only at active AI processors would have produced a much lower figure of around 0.10 Wh per median prompt.
This is an excellent illustration of why AI energy estimates can disagree. Researchers may be drawing the boundary around different parts of the system.
Google also reported that the energy used by its median Gemini text prompt had fallen by a factor of 33 over the preceding 12 months, while its associated carbon footprint had fallen by a factor of 44.
These results were published by Google and had not been independently verified, but they demonstrate how quickly the technology can change.
You can read Google’s full explanation of its measurement methodology for more detail.
A 2026 Study Found a Similar Figure
Independent academic modelling published in the journal Joule in 2026 arrived at a strikingly similar estimate.
Researchers analysing frontier-scale models above 200 billion parameters estimated a median energy use of 0.31 Wh per typical query, with an interquartile range of 0.16 to 0.60 Wh.
The study assumed 500 input tokens and realistic production-scale serving on NVIDIA H100 hardware.
The researchers argued that several widely circulated estimates of AI inference energy were four to 20 times too high because they relied on unrealistic assumptions about how large AI systems are actually served.
One major factor is batching.
A data centre does not necessarily devote an entire high-powered GPU to answering one person’s question. Production systems can process work from multiple users simultaneously, spreading the hardware’s energy demand across many requests.
Utilisation also matters. Simply multiplying the maximum rated power of a GPU by the time required to produce an answer can substantially misrepresent the energy attributable to an individual user.
You can read the 2026 Joule study on AI inference energy for the complete methodology.
So Does an AI Prompt Use About 0.3 Wh?
Sometimes.
The convergence of the OpenAI figure of approximately 0.34 Wh, Google’s measured 0.24 Wh and the 2026 study’s median estimate of 0.31 Wh suggests that a few tenths of a watt-hour is a reasonable ballpark for some ordinary text interactions with efficiently served frontier AI systems.
But that statement needs several qualifications.
It does not mean:
- every AI model uses the same amount of electricity;
- every ChatGPT or Gemini request uses this amount;
- a 20-word answer and a 5,000-word analysis cost the same;
- image, video and audio generation have the same footprint as text;
- all researchers include the same infrastructure in their measurements; or
- the figure will remain accurate as AI systems change.
One of the biggest variables is how much work the model performs before and while producing its answer.
Reasoning Models Can Use Much More Energy
Modern AI systems increasingly use what researchers call test-time scaling.
Instead of immediately generating a short answer, a reasoning model may perform substantially more computation before reaching its final response. Some systems also produce thousands of output tokens, use external tools or perform multiple stages of processing.
The 2026 Joule study estimated that a longer reasoning scenario involving approximately 5,000 output tokens consumed a median of 3.91 Wh per query.
That was around 13 times the median energy of the researchers’ standard query scenario.
The estimated interquartile range was 2.15 to 7.05 Wh, with a significant number of cases exceeding 10 Wh.
This distinction has become increasingly important as AI moves beyond simple chatbots.
An AI agent might:
- search multiple websites;
- read several documents;
- write and execute code;
- call other models or tools;
- analyse the results;
- correct errors; and
- repeat the process until it completes a task.
Calling all of that “one prompt” hides the actual computational workload.
Why Do Some Studies Report 5, 10 or Even 20 Wh?
There is another set of AI energy measurements that produces much larger numbers.
The Stanford AI Index 2026, drawing on model benchmarking research, reported substantial differences between individual AI systems when they were tested with prompts of approximately 1,000 input and 1,000 output tokens.
Some tested models consumed several watt-hours per request. The most energy-intensive examples in the report exceeded 20 Wh.
This does not automatically disprove the approximately 0.3 Wh production figures.
The measurements answer different questions.
The results depend on factors including:
- Model size: larger models may require more computation.
- Architecture: models can activate different amounts of their total parameters.
- Hardware: newer accelerators can perform considerably more work per watt.
- Serving configuration: batching and utilisation can transform per-request efficiency.
- Input length: processing very large documents requires additional computation.
- Output length: producing thousands of tokens usually requires substantially more work than producing a short answer.
- Reasoning effort: some models perform more computation before giving the visible answer.
- Measurement boundary: one estimate may include cooling and idle machines while another measures only GPUs.
This is why quoting a single model benchmark as “the amount of electricity AI uses” is misleading.
Does a Longer AI Prompt Use More Electricity?
Generally, processing more input requires more computation, but the relationship between prompt length and total energy use is not as simple as counting the words you type.
Research on large language models shows that both input and output sequence length can influence energy efficiency, and the relationship may be nonlinear.
However, output length can be particularly important because language models generate their response token by token.
This means asking for a concise answer instead of an unnecessarily long one can sometimes reduce computational work.
It also explains why comparisons based simply on “one AI query” can be misleading.
One query might produce “Paris.” Another might produce a 5,000-word report.
They are both technically one query, but they are not equivalent workloads.
Does Saying “Please” and “Thank You” Waste AI Energy?
This question became surprisingly popular as people began thinking about the environmental cost of AI.
Technically, additional words create additional tokens for a language model to process. So adding unnecessary text is not completely free.
But focusing on a couple of polite words misses the much larger variables.
The choice of model, amount of generated output, reasoning effort, server efficiency and number of interactions are far more important than whether your prompt contains the word “please.”
If being polite helps you communicate clearly or simply feels natural, there is little reason to feel guilty about it.
A more meaningful efficiency strategy is to avoid generating large amounts of information you do not actually need.
How Much Energy Does AI Image Generation Use?
Images are more complicated again.
A 2025 study examining 17 modern image-generation models found an enormous difference in electricity consumption between them: the most energy-intensive models used up to 46 times as much energy as the most efficient models tested.
Image resolution also affected energy consumption, although not consistently across models. Doubling image resolution increased electricity consumption by factors ranging from roughly 1.3 to 4.7 in different systems.
Interestingly, the researchers found that prompt length and prompt content did not have a statistically significant effect on image-generation energy consumption in their experiments.
Typing a more descriptive image prompt was therefore not the major energy issue.
The model and generation settings mattered much more.
The study also found that high image quality did not always require higher energy consumption. Some models simultaneously delivered strong image quality and relatively good energy efficiency.
You can read the research paper, The Hidden Cost of an Image: Quantifying the Energy Consumption of AI Image Generation, for the full results.
Text, Image and Reasoning AI Should Not Be Treated as the Same Thing
This is one of the easiest mistakes to make when discussing AI’s environmental footprint.
“Generative AI” is an umbrella term covering very different workloads.
A useful comparison is not:
AI versus no AI.
It is:
What computational task is being performed, by which system, at what scale?
A small classifier detecting spam, a chatbot answering a sentence, an image generator creating a photograph and an AI agent spending several minutes researching a problem may all use artificial intelligence, but they have very different resource requirements.
What About Training an AI Model?
Most of the figures in this article concern inference.
Inference is what happens after a model has been trained, when you ask it to generate or analyse something.
Training is the much larger one-off computational process used to create the model.
Training large frontier models can require substantial electricity, specialised hardware and lengthy computing runs. However, comparing training energy directly with the energy used by an individual prompt can be misleading.
If a model is used billions of times, inference can eventually account for a very large — and potentially larger — share of its lifetime electricity consumption.
The Stanford AI Index notes that inference is becoming increasingly important as deployed models serve enormous numbers of users.
That means improving the efficiency of everyday AI use can matter just as much as making model training more efficient.
The Bigger Issue Is Scale
A single ordinary AI text query may consume relatively little electricity.
Billions of them do not.
The 2026 Joule study estimated that serving one billion typical queries per day under its baseline scenario would require about 0.7 gigawatt-hours of electricity per day.
If only 10% of those requests became long reasoning queries, the researchers estimated demand could rise to around 1.7 GWh per day.
This is the central paradox of AI efficiency.
Each individual computation can become dramatically more efficient while total electricity consumption still rises because people perform vastly more computations.
Economists sometimes describe this phenomenon as a rebound effect: efficiency lowers the cost of using something, which can encourage people to use more of it.
AI Is Part of a Rapidly Growing Data-Centre System
AI does not run in isolation. It is part of the much larger global data-centre industry.
The International Energy Agency estimates that data centres consumed approximately 485 TWh of electricity globally in 2025.
The IEA expects that figure to roughly double to around 950 TWh by 2030, equivalent to about 3% of global electricity demand.
AI-focused data centres are expected to grow much faster, with their electricity consumption projected to triple between 2025 and 2030.
At the same time, the energy required for an individual AI task is falling rapidly as chips, models and serving systems become more efficient.
The IEA describes AI efficiency improvements as occurring at an unusually fast pace, while warning that growing adoption and increasingly energy-intensive applications are likely to outweigh some of those gains.
For more background on the infrastructure behind these systems, see our beginner-friendly guide to data centre energy use.
Does AI Use More Energy Than a Google Search?
You may have seen claims that an AI query requires five times, ten times or even more electricity than a conventional web search.
This comparison should now be treated with caution.
There are two problems.
First, widely circulated figures for the electricity consumed by a Google search often come from much older estimates, while both search infrastructure and AI systems have changed enormously.
Second, neither “search” nor “AI prompt” describes a standard workload.
A search engine may now use machine learning or generative AI internally. An AI assistant may sometimes answer from a small model and at other times initiate a complex reasoning process.
The 2026 Joule researchers concluded that energy use from efficiently served AI inference is already approaching the same broad scale as mature services such as web search.
That does not mean AI has no energy impact. It means simple multiples such as “AI uses ten times more electricity than search” are increasingly poor descriptions of modern systems.
Why Carbon Emissions Are Not the Same as Electricity Use
Electricity consumption and carbon emissions are related, but they are not interchangeable.
Two identical AI requests can use the same amount of electricity while producing different emissions depending on where and when the computation occurs.
A data centre supplied by a relatively low-carbon electricity grid will generally produce fewer operational emissions per kilowatt-hour than one supplied heavily by coal or gas.
Renewable procurement, nuclear generation, battery storage and the timing of flexible workloads can therefore affect AI’s carbon footprint even when its computational energy requirement remains unchanged.
This is one reason green software engineering considers carbon awareness alongside pure energy efficiency.
Energy Is Only Part of AI’s Environmental Footprint
Electricity is the easiest impact to discuss, but it is not the whole picture.
AI infrastructure also requires:
- water for some data-centre cooling systems;
- semiconductor manufacturing;
- metals and other raw materials;
- construction of data centres and power infrastructure;
- manufacturing of GPUs, CPUs, memory and networking equipment; and
- eventual replacement and disposal of hardware.
This is sometimes called embodied impact: the environmental cost associated with producing physical equipment before it even begins operating.
Measuring only the electricity used while a model generates an answer therefore captures just one part of its lifecycle footprint.
Why We Still Cannot Give One Perfect Number
The biggest obstacle is transparency.
AI companies operate proprietary models on proprietary infrastructure and generally release only limited information about real-world energy use.
Researchers attempting to estimate consumption may not know:
- the exact model architecture;
- how many parameters activate during a request;
- which chips are being used;
- how many users share those chips;
- the average input and output length;
- how much hardware is idle;
- the data centre’s cooling overhead;
- where the computation occurs; or
- the carbon intensity of the electricity.
Even excellent research therefore needs assumptions.
More standardised disclosure from AI providers would make comparisons considerably more meaningful.
What Can an Individual AI User Actually Do?
You do not need to stop using AI to use it thoughtfully.
Some practical habits make more sense than obsessing over every prompt.
Use the Right Tool for the Task
If you already know where a piece of information is located, opening the original source may be more efficient and more reliable than asking an AI system to find and summarise it.
On the other hand, AI may save substantial work when it replaces several searches, helps analyse complicated information or automates a genuinely useful task.
Ask for the Amount of Output You Need
If three bullet points will answer your question, you probably do not need a 3,000-word essay.
Generated output can be an important contributor to inference energy use, especially with reasoning models.
Avoid Regenerating Content Without a Reason
Generating ten images when you only need one useful illustration creates more computational work than generating one or two carefully specified options.
Use Smaller Models When They Are Good Enough
Many services now offer models designed for faster, cheaper tasks. A frontier reasoning model is unnecessary for spelling corrections, simple summaries or routine classification.
Do Not Feel Guilty About Useful Computing
The goal of sustainable technology is not to eliminate electricity use.
Electricity can support education, accessibility, creativity, communication, scientific research and productive work.
The more useful question is whether the resources being consumed are producing enough value to justify their environmental cost — and whether the same result could reasonably be achieved with less.
For broader ideas, see our guide to sustainable internet habits.
The Most Accurate Answer: It Depends
So how much energy does AI actually use?
For an ordinary text interaction with a modern, efficiently served AI system, current production evidence suggests that energy consumption can fall in the neighbourhood of a few tenths of a watt-hour per prompt.
But that number can rise dramatically for longer outputs, larger models, reasoning systems, agents and other computationally intensive tasks.
Image generation varies enormously between models. Video generation introduces another level of computational demand. Training frontier models is an entirely different scale of activity.
And even as individual requests become more efficient, the world’s total AI electricity demand can continue growing because billions more requests are being made.
That is why the best answer is not one viral number.
It is a range, a methodology and a clear explanation of what was actually measured.
As new measurements become available, this guide will be updated to reflect the best current evidence.
Frequently Asked Questions
How much electricity does one ChatGPT question use?
OpenAI CEO Sam Altman stated in 2025 that an average ChatGPT query used approximately 0.34 watt-hours of electricity. OpenAI did not publish a detailed methodology with the figure, so the energy used by an individual request may differ substantially.
How much energy does Gemini use?
Google measured the median Gemini Apps text prompt at 0.24 Wh in May 2025. Its methodology included AI processors, CPUs, memory, idle capacity and data-centre overhead.
Does an AI reasoning model use more electricity?
It can. A 2026 study estimated a median of 3.91 Wh for a long reasoning scenario involving approximately 5,000 output tokens, around 13 times its estimate for a standard frontier-model query.
Does AI image generation use more energy than text?
The answer depends heavily on the model. Image-generation research has found very large differences between systems, including up to a 46-fold gap in energy use among 17 models tested in one 2025 study. It is therefore unreliable to assign one energy figure to every AI-generated image.
Is AI worse for the environment than Google Search?
There is no reliable universal multiplier. Older claims that AI uses many times more energy than web search often combine estimates created using different methodologies and from different generations of technology. Modern AI and search systems also increasingly overlap.
Will AI become more energy efficient?
It already is. Hardware, model architecture and production-serving improvements are reducing the energy required for individual AI tasks rapidly. However, total AI electricity consumption may still increase because usage and the complexity of AI workloads are growing even faster.