How Much Energy Does an AI Query Use? – Per-Prompt Estimates

You type a question, hit enter and get an answer a second later. Somewhere a chip spun up, a fan sped up and a meter ticked forward. For a while the internet repeated a single number for that moment, about 3 watt-hours, roughly ten times a web search. Newer estimates came in far lower, and some came from the companies running the models. Understanding AI query energy use means knowing where those numbers come from and what they leave out. How much energy does one AI query really use?

Recent published estimates put a typical text prompt to a large AI chatbot at roughly 0.2 to 0.4 watt-hours (Wh), well below the older, widely repeated figure of about 3 Wh. Google reported a median of 0.24 Wh for a Gemini text prompt in 2025, and the research group Epoch AI estimated about 0.3 Wh for a typical GPT-4o query. Long inputs, long reasoning answers, images and video can use many times more. The estimates differ mainly in model, prompt length and what they count.

What Are the Published Estimates for AI Query Energy Use?

There are only a handful of serious public estimates, and they do not all measure the same thing. Here are the main ones:

• Google (August 2025): a median Gemini apps text prompt used about 0.24 Wh, with about 0.03 grams of CO2-equivalent and 0.26 milliliters of water, counting active chips, idle capacity, host systems and data center overhead.
• Epoch AI (February 2025): about 0.3 Wh for a typical GPT-4o query, roughly ten times lower than the oft-cited 3 Wh, and 2.5 to 40 Wh for very long inputs.
• OpenAI CEO Sam Altman (June 2025 blog post): an average ChatGPT query uses about 0.34 Wh.
• The older figure of roughly 3 Wh per chatbot query, which circulated widely from 2023 and was built from earlier hardware and usage assumptions.

The chart below lines them up. Note that these come from different methods and companies, so treat it as a comparison of published claims rather than a controlled test.

Why Do Per-Prompt Estimates Differ So Much?

Mostly because they draw the box in different places. A per-prompt number depends on choices about what to count, and small changes in those choices move the answer several times over.

What Counts Inside the Boundary

Some estimates count only the accelerator chips actively working on your prompt. Others add the host processors and memory in the server, idle machines kept ready for demand spikes, and the cooling and power losses captured by power usage effectiveness (PUE). Google's figure deliberately includes those extras, and the company noted that a narrower, chip-only method would give a much smaller number.

Model Size and Hardware

Bigger models do more math per word. Newer chips do that math with less energy. An estimate built on an older model running on older hardware can easily come out ten times higher than one built on a newer, optimized setup.

Prompt and Answer Length

Energy scales with the amount of text processed and generated. A one-line question with a short answer is cheap. Pasting in a long document, or asking a reasoning model that writes out long internal steps before answering, can multiply the energy many times. That is why Epoch's long-input range runs from 2.5 to 40 Wh.

Median Versus Average

A median prompt is the middle of the pack. An average can be pulled up by a small number of very heavy requests. Google reported a median; Altman described an average. They are not the same statistic, even before you account for different models.

What About Images, Video and Reasoning Models?

They use more, often much more, but good public numbers are scarce. Generating an image takes far more computation than producing a short paragraph of text, and video multiplies that again because every second contains many frames. Reasoning models, which work through long chains of intermediate steps before answering, can generate many times more text internally than they show you.

The published figures above are almost all for text prompts. Be careful with any claim that applies a text-prompt number to images or video, or the reverse. The best anyone can say for now is that heavier requests use multiples of the text figure, and the exact multiple depends on the model, the resolution or length, and the hardware.

Does the Per-Prompt Number Include Training?

Generally, no. Per-prompt estimates describe serving a model that already exists, which is called inference. The energy used to train the model in the first place is a separate, one-time cost spread over every prompt the model ever answers.

That makes training harder to allocate. A heavily used model spreads its training energy across billions of prompts, so the training share per prompt shrinks. A model that is trained and then rarely used carries a much larger share per answer. Most published per-prompt figures leave training out and say so, which is reasonable as long as readers know it.

Google's 2025 report also put numbers on the other side effects of a median text prompt: about 0.03 grams of carbon dioxide equivalent and 0.26 milliliters of water, roughly five drops. Those figures depend on the company's power contracts and cooling choices, so they vary even more between operators than the energy figure does.

How Big Is 0.3 Watt-Hours, Really?

It is small for one person. Three-tenths of a watt-hour is about what a 10-watt LED bulb uses in under two minutes: 0.3 Wh ÷ 10 W = 0.03 hours, or 1.8 minutes. Even at 3 Wh, a single query would run that bulb for 18 minutes.

For comparison, Google said back in 2009 that a typical web search used about 0.3 Wh. So the newest chatbot estimates are in the same ballpark as that old search figure, though search itself has changed a great deal since then and the two are not directly comparable.

The point is not that a prompt is free. It is that the energy question lives at scale, not at the keyboard.

What Happens When You Multiply by Billions?

The numbers get big quickly. Here is an illustrative example using round inputs, not any company's reported traffic:

1. Assume 1 billion text prompts per day at 0.3 Wh each.
2. Daily energy: 1,000,000,000 × 0.3 Wh = 300,000,000 Wh, or 300 megawatt-hours (MWh).
3. Average power: 300 MWh ÷ 24 hours = 12.5 megawatts (MW), running around the clock.
4. Yearly energy: 300 MWh × 365 = 109,500 MWh.

Now change one input. If the average prompt uses 3 Wh because of longer answers, documents or reasoning, the same traffic needs 125 MW. Add image and video generation and the total climbs again. This is why the shift toward longer, more complex requests matters as much as the raw count.

At that scale, serving prompts becomes a power plant decision. Operators weigh grid connections, on-site generation and hybrid designs, and some look at a behind-the-meter data center to get firm power on a faster schedule than the grid can offer.

Is the Energy per Prompt Going Down?

For text prompts, yes, and fast. Google reported that energy per median Gemini text prompt fell 33-fold over twelve months, crediting more efficient models, better software and hardware, and its data center fleet. Newer chips, batching many requests together and routing simple questions to smaller models all push the per-prompt number down.

Total energy is a different story. As each answer gets cheaper, people ask more questions, request longer answers and use AI in more products. Lower energy per prompt and higher total energy can happen at the same time, and so far they have.

Reading Any AI Query Energy Use Number

If you remember one thing about AI query energy use, make it this: the per-prompt number is small and falling, and the total is large and rising. Both facts matter, and anyone who quotes only one of them is telling half the story.

When you see a new number, ask four questions: which model, how long was the prompt and answer, is it a median or an average, and does the boundary include idle capacity and cooling? The answers will tell you whether the figure is comparable to anything else, and whether it says more about one prompt or about the power plant behind a billion of them.

← All guides