A median text prompt to Google's Gemini uses about 0.24 watt-hours of electricity — roughly what a 60-watt bulb burns in fifteen seconds. That figure comes from Google's own production measurement of its live service, published in August 2025.
It is about one-twelfth of the number most news coverage still repeats.
The widely quoted figure is 3 watt-hours per query, usually phrased as "ten times a Google search." It was never measured. Follow it back and it ends at two places: a Google blog post from 2009, and a single remark in a 2023 interview that was not about energy at all.
That gap matters, but not for the reason you might expect. The useful conclusion is not "AI is cleaner than you thought." It is that the per-query number is the wrong number to argue about — and once you understand why, the energy debate gets considerably clearer.
Where the "Ten Times a Google Search" Number Came From
In February 2023, Reuters interviewed John Hennessy, the chairman of Alphabet, about what generative AI would cost the company. He said that an exchange with a large language model "likely cost 10 times more than a standard keyword search."
Read that again: cost. Hennessy was talking about dollars — chips, servers, engineers, margin. He was not talking about watt-hours, and he had no connection to OpenAI or Microsoft, whose products the figure would later be attached to.
Later in 2023, that ratio was converted into an energy estimate of roughly 3 Wh per interaction. To make the conversion you need a baseline for a search, and the only official one Google has ever published appeared on its corporate blog in 2009: "this amounts to 0.0003 kWh of energy per search," or 0.3 Wh.
So a 2023 comment about money was multiplied by a 2009 figure about a search engine that predates transformers entirely, and the result became the standard fact about AI's energy use.
It spread. Researchers at Hugging Face, Salesforce, Ekimetrics and Carnegie Mellon sampled 100 news articles about ChatGPT's energy consumption published up to April 2025. 53% of them cited the 3 Wh figure or the "ten times a Google search" claim. Their paper, Misinformation by Omission, argues the real failure is upstream: when companies publish nothing, estimates fill the vacuum and then harden into facts.
This is a pattern worth recognising in its own right, and not only here — we have written about how AI benchmark numbers travel further than the evidence behind them for much the same reason.
What the Measured Numbers Actually Say
Three independent figures now exist, and they agree with each other far more than any of them agrees with 3 Wh.
Four Numbers for One AI Query
Each measures a different thing. Only one of them is a production measurement.
| Figure | Source | What it actually is |
|---|---|---|
| Figure~3 Wh | Source2023 estimate, still widely quoted | What it actually isA cost remark converted using a 2009 search figure |
| Figure0.24 Wh | SourceGoogle, August 2025 | What it actually isMedian Gemini Apps text prompt, measured in production |
| Figure~0.3 Wh | SourceEpoch AI, February 2025 | What it actually isCalculated for GPT-4o from chip time and utilisation |
| Figure~0.34 Wh | SourceOpenAI, 2025 | What it actually isAn average query, stated without published method |
Three methods, one order of magnitude
A production measurement, an independent calculation and a vendor statement all land between 0.24 and 0.34 Wh. The outlier is the number everyone quotes.
Google's number is the most useful of the three because it is the only one with a published methodology and real traffic behind it. It describes the median text prompt, using May 2025 data, and comes with two companion figures: 0.03 grams of CO₂ equivalent and 0.26 millilitres of water — about five drops.
Two honest caveats belong with it. Google states plainly in its own footnotes that "the data and claims have not been verified by an independent third-party," and the figures cover text prompts only. Image and video generation were not included and cost substantially more — and how AI video generation works explains why the per-output cost is so much higher.
The Boundary Does More Work Than the Number
Here is the part that gets skipped, and it is the single most important thing in this article.
Google published two figures for the same prompt. The headline 0.24 Wh is the comprehensive one. A narrower accounting — counting only the energy drawn by the TPU and GPU chips while they are actively working on your prompt — gives 0.10 Wh.
Same prompt. Same model. Same day. The number is 2.4 times larger depending only on where you draw the line.
What the wider boundary adds is not padding. It is idle machines held in reserve so the service answers instantly, the host CPU and memory feeding the accelerators, and the cooling and power-conversion overhead of the building itself. Google calls the narrow approach "an optimistic scenario at best" that "substantially underestimates the real operational footprint of AI" — an unusually direct thing for a company to say about its own lower number.
And all four figures in the table share the same exclusions. None of them counts training the model, the network carrying the response to you, or your own device displaying it. Training is a separate, very large, one-off cost that is then amortised across every query the model ever answers — which is why it belongs in a different calculation rather than this one. We cover that side in how large language models are trained.
So before comparing any two AI energy figures, check that they are drawing the same line. Usually they are not.
The Median Hides the Distribution
A median is the middle of a range, not a ceiling, and the range here is wide.
Epoch AI's analysis is useful precisely because it shows its working. Its ~0.3 Wh estimate assumes a typical short exchange of roughly 500 output tokens. Feed the same model a 10,000-token input — a long article, a contract, a chunk of a codebase — and by the same method the query costs around 2.5 Wh, some eight times more. If you want the mechanics of why longer inputs cost more, that is the subject of what a context window is and why it matters.
Three things push a query up the distribution:
- Long inputs. Everything you paste in has to be processed before a single word comes back.
- Long outputs. Energy scales with tokens generated, and reasoning models deliberately generate thousands of intermediate tokens before producing an answer.
- Different modalities. Image and video generation are not in Google's 0.24 Wh figure at all.
A short factual question and a request to analyse a fifty-page document are not the same workload, and no single number describes both.
Work Out Your Own Number
The arithmetic is simple enough to do yourself, which is the best defence against a quoted figure.
Say you send 20 text prompts a day, every day, at the measured median:
- 20 × 0.24 Wh = 4.8 Wh per day
- 4.8 Wh × 365 = 1,752 Wh, or about 1.75 kWh a year
For scale, the average US residential utility customer used about 865 kWh a month in 2024, according to the Energy Information Administration — roughly 10,400 kWh a year. A year of twenty daily prompts is about 0.02% of that. Put differently, it is a 60-watt bulb left on for about 29 hours.
Now run it again assuming every one of those prompts carries a long document, at Epoch's 2.5 Wh figure: 18.3 kWh a year, around ten times more, and still well under 0.2% of a household's draw.
Both results point the same way. Individual use is not where this story is. Which brings us to the number that does matter.
The Totals Rise Anyway
Google's own report contains a striking claim: over twelve months, the energy footprint of its median text prompt fell by a factor of 33, and its carbon footprint by a factor of 44, while response quality improved.
Per-query efficiency improved by more than an order of magnitude in a single year. Total data-center electricity demand rose anyway.
- The International Energy Agency puts global data-center consumption at about 415 TWh in 2024, roughly 1.5% of world electricity, and projects about 945 TWh by 2030 in its base case — more than double, and comparable to Japan's entire national consumption today.
- Lawrence Berkeley National Laboratory, in the report Congress required under the Energy Act of 2020, put US data centers at 176 TWh in 2023 — 4.4% of national electricity — rising to between 325 and 580 TWh by 2028, or 6.7% to 12%.
There is no contradiction between those two facts, and it is the central tension running through everything we publish on climate and energy. Efficiency per query sets the cost of one unit; it does not cap how many units get requested. When a thing gets cheaper and better at the same time, people use far more of it, and the total can climb while every individual transaction gets leaner.
This is why the per-query debate is a distraction. Nothing you decide about your own prompting changes a 945 TWh projection. What changes it is how fast data centers are built, what generates the electricity they draw, and whether the grid can connect them at all — which is a separate and much harder problem, covered in why AI data centers are straining the power grid.
How to Read Any AI Energy Claim
Four questions will tell you how much weight a figure deserves. They work on vendor reports and critical coverage equally.
- Is it a measurement or an estimate? Production telemetry, an outside calculation, and a press remark are three very different kinds of claim, and all three get quoted in the same sentence.
- Where is the boundary? Chips only, or chips plus idle capacity plus building overhead? That choice alone moved Google's figure by 2.4 times.
- Median, average, or a specific prompt? And of what — text, image, video, a reasoning model?
- Per query, or in total? Almost every argument that sounds like a disagreement is actually two people answering different questions.
A figure that survives all four is worth repeating. The 3 Wh figure fails the first one.
The Bottom Line
The best available measurement says a median AI text prompt uses about 0.24 watt-hours — roughly a twelfth of the figure in wide circulation, which was never a measurement and was originally about cost rather than energy. That correction is worth making, because arguments built on bad numbers tend to produce bad decisions.
But the correction does not mean the concern was misplaced. It means the concern was aimed at the wrong level. Per-query energy is small, falling fast, and almost entirely outside your control. Aggregate demand is large, rising fast, and decided by infrastructure choices being made right now. The honest summary is that your prompts are not the problem, and the data centers answering them still are.
Frequently Asked Questions
Does one ChatGPT query really use ten times more energy than a Google search?
There is no good evidence for it. The claim traces to a February 2023 Reuters interview in which Alphabet chairman John Hennessy said an LLM exchange likely cost ten times more than a keyword search — a statement about money, not electricity. It was later converted into an energy figure using a Google search baseline published in 2009. The measured figures available since 2025 put a median AI text prompt at 0.24 to 0.34 Wh, close to the old search estimate rather than ten times it.
How much water does an AI query use?
Google measured 0.26 millilitres — about five drops — for the median Gemini text prompt, alongside 0.24 Wh of energy. That covers data-center cooling within the same boundary as its energy figure. It excludes the water used to generate the electricity itself, which varies enormously depending on whether the power comes from hydro, nuclear, gas or solar.
Do reasoning models use more energy than regular ones?
Yes, and the difference is structural rather than incidental. Reasoning models generate long chains of intermediate tokens before producing an answer, and energy use scales with tokens generated. Epoch AI estimates that queries with long inputs or long outputs can be several times more expensive than a short exchange — around 2.5 Wh for a 10,000-token input against roughly 0.3 Wh for a short one.
Is training or running a model more energy-intensive?
Training is far larger as a single event; inference is far larger in aggregate over a popular model's lifetime, because it happens billions of times. The per-query figures discussed here exclude training entirely. Any comparison between the two has to state how many queries the training cost is being spread across, and most do not.
Should I stop using AI tools to save energy?
The arithmetic does not support it as a climate decision. Twenty text prompts a day for a year comes to about 1.75 kWh — roughly 0.02% of an average US household's annual electricity use, or a 60-watt bulb left on for a little over a day. Heating, driving and flying are orders of magnitude larger. The meaningful decisions about AI's energy footprint are made in procurement and grid planning, not in individual prompting.
Why don't we have better numbers than this?
Because disclosure is voluntary and mostly absent. Google's August 2025 report was the first production measurement of its kind from a major provider, and it notes that its own data has not been independently verified. OpenAI has given a figure without a published method. Most providers publish nothing at all, which is the central argument of the Misinformation by Omission paper: the estimates people criticise exist because the measurements do not.
Sources
- Measuring the environmental impact of AI inference — Google Cloud, August 2025
- Misinformation by Omission: The Need for More Environmental Transparency in AI — Luccioni, Gamazaychikov, Alves da Costa and Strubell, arXiv, June 2025
- How much energy does ChatGPT use? — Epoch AI, February 2025
- Energy and AI — International Energy Agency, April 2025
- 2024 United States Data Center Energy Usage Report — Lawrence Berkeley National Laboratory, December 2024
- How much electricity does an American home use? — US Energy Information Administration
