Wired traces the data center buildout to agents that prompt themselves for hours

A user types one sentence asking for a website and closes the laptop. According to Maxwell Zeff, who writes Wired's Model Behavior newsletter, what runs after that is not a single inference. "Rather than asking an AI chatbot a simple question and answer, these agents can give themselves hundreds of small prompts based on a user's original question," he said ^[4]. An agent given the website job "might run for hours to build out features, re-prompting itself dozens of times in the process to build different web pages, menus, and datasets that power the thing," he said ^[5].

Boris Gamazaychikov, co-founder and CEO of the research and advisory group Sustainable AI, told Taft the constraint has changed. "In other technological growth areas, we're constrained by how many people are driving a car or streaming Netflix," he said ^[12]. He said AI leaders are "talking about unicorns that have one employee" ^[13]. Taft writes that the work is decoupled from users ^[14]. That removes the assumption every seat model rests on: demand scales with humans at keyboards.

OpenAI's math run shows the top end. The company announced that a swarm of more than 10,000 agents sending 2.7 million messages solved a longstanding math problem ^[6]. That is an average of 270 messages per agent ^[16]. Zeff estimated the energy at probably tens of millions of dollars and said how much exactly is tough to say ^[8]. Take the bottom of that range, ten million dollars, and assume all of it went into the run: each message then carries about $3.70 of energy ^[17]. Mathematicians pushed back on the company's claims ^[7], and Taft describes the run as an outlier ^[19].

Private AI companies have historically been choosy about what they disclose on environmental metrics, Taft writes, and many CEOs point to single queries made by individuals as a measure of resource use ^[9]. Sam Altman, in a recent podcast interview, claimed the water needed to harvest a single almond amounted to 38,000 ChatGPT queries, a calculation that has been disputed ^[10][11]. "The people that are scarfing down 12 almonds at a time don't feel like they're doing something horrible from a water perspective for the most part," he said ^[15]. On his own figure, twelve almonds is 456,000 queries ^[18]. The unit there is the individual query, and Taft reports agents are much more energy-intensive than simple queries ^[20].

Wired reports a major dearth of information on agent energy use, with tasks that run from simple jobs up to a full day of autonomous coding involving a team of parallel helper agents, and a massive gulf in power use between those cases ^[21]. The people quoted assert the per-task multiplier; no one has measured it. Climate scientist Zeke Hausfather published a blog post last month working out the energy of his own AI use, which leans heavily on agents, using a variety of different sources ^[22].

Attended work with a step count you set behaves the way chatbot-era planning assumed, and per-seat numbers still track it. Unattended work whose step count the model sets is where seat counts stop predicting spend. For anything in that second group, the two figures to log per task are wall-clock runtime and the number of times the task prompted itself, and both come out of your own traces and vendor bills.