From Futurism · Joe Wilkins · · 1m 30s
Your AI coding agent eats power like two fridges
A climate scientist measured his own Claude Code habit and found a daily bill most users never see.
In brief
A climate scientist measured his own Claude Code habit and found a daily bill most users never see. Agents burn most of their energy re-reading their own context, not producing the text you actually read. Originally reported by Futurism.
Agents changed the question
Measuring one chatbot prompt no longer tells you much. Modern agents stack a language model with tools and app layers that each pull power to finish a task on their own.
The labs stay quiet
Top private AI companies publish almost nothing about internal energy use, so nobody can say for sure what AI assistants are doing to the grid.
One scientist ran the numbers
Climate scientist Zeke Hausfather counted the tokens his Claude Code sessions consumed and pegged his average day at 1.2 to 5.9 kWh, more than two refrigerators running all day.
Re-reading its own memory
Over eight weeks he burned 3.2 billion tokens. Most went to the agent re-processing its whole accumulated context at every single step it took.
“Every time an agent takes a step, the model re-processes its entire accumulated context”
You see 0.4 percent
Across roughly 14,000 steps, 96 percent of tokens were cache reads and only a sliver was visible output. Cached tokens are cheaper, which is why the estimate is a range.
“The text I actually see is only around 0.4 percent of total tokens processed”
Small per person, new in total
A year of this is about one electric clothes dryer. Spread across every AI user it becomes a fresh source of emissions, and a researcher says better data lands this month.
“represents a net new source of emissions”
The key point
Agents burn most of their energy re-reading their own context, not producing the text you actually read.
