Climate scientist Zeke Hausfather tallied the tokens his own Claude Code sessions consumed over eight weeks and put the daily electricity draw at 1.2 to 5.9 kWh — for a single user, more than two refrigerators running around the clock. The estimate, flagged by Wired's Molly Taft, exists because nobody better positioned will make it. The large private labs disclose almost nothing about their own energy consumption, which leaves the effect of AI agents on the grid unmeasurable from the outside. Hausfather measured it from inside his own terminal instead.
The breakdown is the interesting part. Over those eight weeks the agent processed 3.2 billion tokens across 14,000 tool steps. About 96% of them were cache reads — the model re-loading its own accumulated working memory every time it ran a command, opened a file or called a tool. The text Hausfather actually saw, the model's visible output, came to roughly 0.4% of the total: about one token in 250. Almost all the compute went into the agent reminding itself what it was doing.
That ratio breaks the unit everyone has been arguing about. Energy per prompt assumes a prompt is a discrete transaction: text goes in, an answer comes out, the meter stops. An agent has no such boundary. It reprocesses everything it has accumulated at each new step, and on Hausfather's own totals the average step moved something close to 230,000 tokens. What the user types is not the workload. It is the trigger for the workload.
The cache share cuts both ways. Reprocessing a cached token is cheaper than processing a fresh one, so 3.2 billion tokens do not convert into 3.2 billion tokens' worth of electricity — the overwhelming majority of that traffic runs at a discount, which is exactly why the accounting is hard. It is also why the result arrives as a range rather than a figure. A spread of 1.2 to 5.9 kWh is nearly five to one, and that reads less like a measurement than an admission of how little is known.
Scaled up to a year, the total lands somewhere modest. Hausfather estimates the emissions as slightly more than running an electric clothes dryer, and about half those of driving an electric car 11,500 miles in California. A dryer and a long commute are the honest comparisons for one person's usage, and they are also why this argument keeps stalling: no individual's number is alarming. Multiplied across a user base it adds up quickly, and it adds up as a new source of emissions at a moment when global temperatures are rising and emissions targets keep sliding further out.
My reading is that the energy story here is downstream of an architecture story. Ninety-six percent cache reads is not a law of physics. It is a description of how agents are currently built — models with no persistent memory handed the entire transcript again at every step because there is no other way for them to know what happened. If that changes, through better memory, smaller working contexts or cheaper cache serving, the consumption curve moves substantially without anyone using the tool any less. The more interesting question is not what a coding agent burns today. It is whether that 96% is a permanent property of the design or a temporary one.
Hausfather's arithmetic already has a critic. As Wired reports, Boris Gamazaychikov, who researches the energy consumption of AI agents, called the calculation a good-faith attempt but said it rests on outdated data. His own study on agent energy use is due later in September, and it could shift the picture of what these systems actually cost.
So the best public number on agent electricity use is a five-to-one range, produced by a customer, disputed for using stale inputs, and awaiting replacement by research that has not appeared yet. The companies that could settle the question are the ones running the hardware, and they have chosen not to. Until that changes, every estimate of what agents draw from the grid will be a reconstruction from the outside — and the distance between the estimates will stay wider than the quantity they are trying to measure.