In the age of AI agents, it’s no longer enough to ask how much electricity a single chatbot prompt uses. Today’s AI models are even more complicated systems, often designed to complete complex tasks autonomously, and are made up not only of large language models, but a host of other tools and application layers which demand much more power than the humble standalone chatbot.

That makes calculating their total electricity usage a much bigger can of worms. Yet with virtual radio silence from the top private AI labs on internal energy use, it’s nearly impossible to say definitively what kind of impact our AI assistants are having on the electrical grid.

In a blog post flagged by Wired‘s Molly Taft, the climate scientist Zeke Hausfather took it upon himself to calculate the likely energy demand behind his typical sessions with Claude Code, an AI-coding assistant developed by Anthropic.

Measuring the number of tokens his AI assistant consumed, Hausfather estimated that his average daily use fell in the range of 1.2 to 5.9 kWh — an alarming amount for a single consumer, consuming more electricity than two refrigerators running all day long.

Most of this token consumption, 3.2 billion over an eight-week period, came from the agent “re-reading its own working memory,” Hausfather wrote, which happens every time the tool takes a new step.

“Every time an agent takes a step (eg, runs a command, reads a file, or calls a tool), the model re-processes its entire accumulated context,” the climate scientist wrote.

“The text I actually see (eg, the model’s output) is only around 0.4 percent of total tokens processed,” Hausfather continued. “Some 96 percent of the tokens are cache reads where the agent re-reads its own context at each of those 14,000 steps. This matters enormously for the energy estimate, because a cached token is much cheaper to re-read than a fresh one is to process.”

Extrapolating those results out over a year, Hausfather contends that he would emit “a bit more per year than running an electric clothes dryer, and about half as much as driving an electric car 11,500 miles in California.”

That might not sound like much, but over an entire population of AI users, it adds up quickly. Likewise, it “represents a net new source of emissions, at a time when global temperatures are skyrocketing and our emissions reduction goals are increasingly off track,” the climate scientist explains.

The exact climate toll, though, remains hazy. As Wired notes, AI sustainability researcher Boris Gamazaychikov called the calculation a “good effort,” but one based on old data. Gamazaychikov’s own research on agentic AI energy consumption should be due out later in September — and when it drops, it could drastically change our understanding of the true cost of AI agents.

More on AI: Sam Altman Now Trying to Gain Control of Electric Grid