AI agents push energy use far beyond the per-query math
Wired reports that agents re-prompting themselves hundreds of times per task are a major driver behind the current data center buildout.
Wired's data center column Power Play argues that the rush to build power plants and data centers is driven by agents rather than ordinary chatbot use. Agents are large language model (LLM) systems that make autonomous decisions to carry out a task, generating hundreds of small prompts for themselves from a single user request.
Wired writer Maxwell Zeff gives an example: asked to build a website, an agent may run for hours, re-prompting itself dozens of times to produce pages, menus and the datasets behind them.
Few numbers, many estimates
OpenAI recently announced that a swarm of more than 10,000 agents exchanging 2.7 million messages had solved a longstanding math problem; mathematicians pushed back on the claim. According to Zeff, the processing involved probably cost tens of millions of dollars in energy, though an exact figure is hard to pin down.
Companies disclose little about environmental metrics. On a recent podcast, OpenAI CEO Sam Altman said the water needed to grow a single almond was equivalent to 38,000 ChatGPT queries; that calculation has been disputed.
Boris Gamazaychikov, co-founder and CEO of the research group Sustainable AI, says consumption in areas like driving or streaming is capped by how many people are doing it, while agent workloads are decoupled from users.
Two refrigerators
Climate scientist Zeke Hausfather published a blog post last month calculating his own usage, which leans heavily on agents, and concluded that his average daily Claude session may consume more energy than it takes to run two refrigerators. Gamazaychikov said the estimate relied on somewhat outdated findings; his group plans to publish more precise figures for agents running on closed models this month.
Meta last week introduced a personal agent called Muse, saying it keeps a dedicated computer in the cloud for each user that works even when the user is offline. The company plans to integrate Muse with its AI glasses later this year.
Full column at Wired.
