Asking an AI assistant a question uses electricity – and I’m not talking about your phone. AI data centres have been in the news more and more recently, with particular scrutiny surrounding their energy use. But are all the stats accurate? Let’s do our best to find out.
Let’s set the scene. You type, wait a few seconds, and read an answer. Meanwhile, computers in a data centre have been doing the work – although there’s no handy meter beside the chat to tell you how much energy they’ve used.
That makes it difficult to judge what an AI search actually uses. Some published figures put an ordinary text response at a few tenths of a watt-hour. But a chatbot answering a simple question and an AI assistant spending several minutes researching your next hot tub purchase aren’t doing the same amount of work.
I recently looked at AI’s wider electricity and water footprint. This time, let’s concentrate on the searching itself – and what happens when you press send.
How much energy does AI use for a quick, basic answer?
One useful starting point is Google’s study of Gemini Apps. It reported that the median text prompt used 0.24 watt-hours (Wh) of electricity in May 2025, including the computers doing the work, spare capacity, and data-centre overhead like cooling.
That gives us a useful reference, but it doesn’t cover every AI service or today’s models. It’s also a median figure, which won’t capture how much unusually demanding requests add to the overall electricity bill.
Separately, a February 2025 analysis published by Epoch AI estimated around 0.3Wh for a typical GPT-4o query. This was a calculation based on assumptions about the model, hardware, and response length, rather than a reading from OpenAI’s electricity meter.
Still, 0.3Wh gives us something understandable to work with. It’s enough to run a 10W LED bulb for about one minute and 48 seconds, if you want to be super specific. If 100 responses each used that amount, they’d consume 30Wh – equivalent to leaving that bulb on for three hours.
You can’t audit your chat history with those sums, but they give a sense of the scale. In other words, you can rest assured that your ordinary text answer hasn’t secretly switched on the equivalent of 1000 kettles.
It’s not always that simple
Before you reach for your environmental halo though, let’s take a step back. Searching the web gives the AI more to do – before answering, it may have to find information, read it, and decide whether it needs to look for anything else.
Google describes how AI Mode uses a technique called query fan-out. It breaks a question into subtopics, runs related searches, and brings the results together. So asking it to compare a few laptops could involve looking up specs, reviews, and other information before it writes its recommendation.
That can be useful, obviously. I’d much rather get an answer that’s checked something than one that’s confidently made it up. The catch, is that we can’t use the Gemini Apps figure above as the electricity cost of an AI Mode search.
Longer reasoning adds another variable. A research paper updated in June 2026 estimated a median of 0.31Wh per query for large models running under its assumed deployment conditions. In scenarios involving 15 times as many tokens – the chunks of text models process – the estimate rose to 3.91Wh –about 13 times as much!
That isn’t a measured price tag for your chatbot’s research button, obviously. But it does show why counting questions alone is a poor guide. A short question can set off a lengthy job, including intermediate reasoning that never appears in the final answer.
How does AI compare with ordinary search?
You may have seen the claim that an AI query uses ten times as much electricity as a conventional Google search. It sounds like a handy rule. Unfortunately, the numbers behind it are showing their age.
The widely repeated 0.3Wh figure for a Google search comes from a Google blog post published in 2009. Yes, 2009. The roughly 3Wh estimate often used for the AI side also relies on older hardware and workload assumptions, which the Epoch analysis challenged.
We can’t use those old figures to compare searches made today. And the newer 0.24Wh Gemini figure doesn’t prove that AI now uses less electricity than ordinary search, either – we’d need up-to-date measurements of both, counting the same things.
There’s also the question of what you’re replacing. Getting one useful answer could save several searches and page visits. Equally, asking an assistant to produce a sprawling report when you only wanted a shop’s opening hours is a rather elaborate way of finding out that it shuts at six.
The electricity that isn’t in the answer
individual prompt figures generally describe the work of running a trained model, known as inference. Google’s energy measurement, for example, leaves out model training and the electricity used by your own device. In other words, it’s far from a complete, accurate lifetime energy bill.
Using less electricity usually helps, but the source matters too. Powering a server with coal-fired electricity has a different carbon footprint from powering it with wind energy. An energy figure alone won’t tell you the emissions.
And small individual requests can still add up to substantial demand. The International Energy Agency’s 2026 outlook puts global data-centre electricity consumption at 485 terawatt-hours in 2025 and projects roughly 950TWh in 2030. A terawatt-hour is a billion kilowatt-hours, so we’ve moved some distance beyond the desk lamp.
Those totals cover all data centres, not just AI. However, the agency expects electricity consumption from AI-focused facilities to triple over that period. More efficient answers don’t automatically mean less electricity used overall if demand grows faster than the savings.
Using the right amount of AI
For everyday questions, I wouldn’t automatically switch on the most intensive research mode. And if you only need a short answer, ask for one. You don’t have to accept an essay just because the chatbot’s keen to provide one. You can also tell it to be less verbose.
Having said all that – if you’re comparing complicated information or researching something properly, the extra work may be worth it. I’m certainly not going to insist on you accepting a worse answer just to save an unknown amount of electricity.
I would, however, like the services to make that choice easier. Show me an approximate energy figure alongside the different modes, with an explanation of what it includes. A rough estimate would be more useful than leaving me to work it out from last year’s research papers. Likely? No. But at least dreaming is free. For now, at least.
Read the full article here
