A climate scientist tracked his own AI use for eight weeks and measured the AI agent power behind a single request at about 150 watt-hours of electricity. That is roughly 600 times what a normal chatbot question costs, and almost none of it happens where you can see it.
The Gist
- One typed request to an AI agent averages 150 watt-hours of electricity.
- A regular chatbot question sits between 0.24 and 0.34 watt-hours.
- Over eight weeks, 1,138 requests triggered more than 14,000 hidden calls to the model.
- 96% of the work was the agent rereading things it had already been told.
Have ChatGPT Recap This Article
ChatGPTMeet the AI agent, the assistant that works in steps
A chatbot answers you once. You type a question, it types an answer, and the exchange is over. That is the version of AI most people have used.
An AI agent is different. You give it a goal instead of a question, something like fix this file or book me a table, and it goes off and does several things in a row on its own before coming back to you.
Those in-between steps are the whole story. Every step is a fresh request to the AI, even though you only typed once, and each one draws power in a data center somewhere.
Agents have been spreading fast into the tools ordinary people already touch. We saw one land right in the middle of the most popular chatbot when the ChatGPT agent started taking over from the chat window itself.
How one typed request turns into fourteen thousand calls
The measurement is unusual because it comes from real daily use, not a lab test. Over eight weeks of working with a coding agent, the researcher typed 1,138 requests by hand.
Those 1,138 requests set off more than 14,000 calls to the model and moved 3.2 billion tokens. A token is just a chunk of text, roughly three quarters of a word, and it is the unit AI companies count and bill.
Here is the part that explains the AI agent power figure. 96% of those tokens were the agent rereading its own notes, going back over the conversation and the files at every single step rather than producing anything new.
Add it up and you get roughly 170 kilowatt-hours over the eight weeks, though the researcher is honest that the real number sits somewhere between 70 and 330. He published the full calculation and every assumption behind it.
Agents also keep gaining new abilities that lengthen the chain further. Giving one the ability to browse, as happened when Claude Code learned to open web pages on your behalf, means another round of rereading for every page it visits.
Keep learning on AI Noobies:
- AI Interview Prep Beats Rereading the Job Ad
- Suno Will Watermark Every Song You Make
- Delete ChatGPT Memory: See and Erase What It Knows
What a watt-hour looks like in your own home
Numbers like 150 watt-hours mean nothing until you put them next to something you own. A watt-hour is the electricity a one-watt device uses in one hour, so 150 of them is a bright old-style bulb left on for a couple of hours.
Stretched across a working day, that AI agent power adds up to about 3 kilowatt-hours. The researcher’s own comparison is blunt: his daily agent habit pulled more power than two fridges running non-stop.
Over a year that works out to roughly 1.1 megawatt-hours and around 370 kilograms of carbon dioxide. For one person, using one tool, on top of everything else they already run.
That said, a plain chatbot question really is tiny. Google puts a typical query at 0.24 watt-hours and OpenAI cites 0.34, which is a second or two of a microwave. What drives the gap is delegation: the moment you stop asking and start handing over a whole task.
How to see the difference in your own AI use this week
You do not need a meter to notice the split. Pay attention to whether your tool answers you straight away or goes quiet for thirty seconds while a little status line ticks through steps. The quiet version is the expensive one.
Most free plans still hand you the cheap kind by default, and that is worth knowing before you upgrade. Free agent access has started appearing too, as it did when Claude Sonnet 5 became free and better at agent work.
The wider argument this opens up is about honesty in the numbers. AI companies publish the per-question figure because it is flattering, and the AI agent power figure is the one nobody advertises.
Expect to hear more of this as data centers keep making local news. When a town argues about a new one, the thing being built for is no longer a billion people asking short questions, it is a smaller group handing over long tasks.
Stay tuned on AI Noobies.
About the Author
