The argument about AI and electricity has been conducted in the wrong unit. Executives answer questions about data centers by pricing a single chatbot query, and the industry's actual product no longer arrives in single queries. AI agents — systems built on large language models that make their own decisions on the way to finishing a task — turn one request into hundreds. That shift, not a user looking up a recipe or a holiday destination, is what the new energy infrastructure is being built for.
Maxwell Zeff, who writes the weekly newsletter Model Behavior, describes the mechanics plainly: instead of one question and one answer, an agent can generate hundreds of small prompts from the user's original request. Ask it to build a website and it may spend hours adding features, rewriting its own instructions dozens of times to assemble the pages, menus and data sets the site needs to run.
The extreme version is already on display. OpenAI said a swarm of more than 10,000 agents, which sent 2.7 million messages between them, solved a long-standing mathematical problem. Mathematicians disputed the company's claims. Zeff estimates the compute behind those messages probably cost tens of millions of dollars, though the exact figure is hard to establish. It was an exceptional case — labs are willing to spend unusual amounts on problems considered unsolvable — but it marks how far the spending can run.
Against that, the industry's preferred yardstick looks strange. Private AI companies decide for themselves which environmental figures to disclose, and executives keep reaching for the single query of a single user. On a recent podcast, OpenAI chief executive Sam Altman said that growing one almond takes as much water as 38,000 ChatGPT queries. The calculation was disputed. Altman added that people who eat 12 almonds at a time mostly do not feel they are doing anything terrible in water terms.
Agents make that comparison much harder to sustain. There is almost no data on what they consume. Their tasks range from the trivial to a full day of autonomous programming with a team of parallel helper agents running alongside, and the gap in energy between those two cases is enormous. As tasks get harder, consumption can grow with almost no ceiling.
Boris Gamazaychikov, co-founder and chief executive of the research and consulting group Sustainable AI, points out that other technology booms had natural limits: the number of people who drive cars, the number who watch Netflix. With AI, the link between the number of users and the volume of work done may disappear entirely. Judging by what AI executives say, he argues, that is precisely the goal — a world in which some companies have a single employee. One person, with hundreds or thousands of agents working in the background. Nobody knows how likely that future is, but it is the one the companies are building toward, and it explains a good deal of the hurry.
Because reliable consumption data is scarce, some people have started counting on their own. Last month the climate scientist Zeke Hausfather published an analysis estimating the energy behind his own AI use, agents included. Drawing on several sources, he concluded that an average daily session with Claude may use more energy than running two refrigerators. Gamazaychikov, whose group is due to publish a study later this month with more precise estimates of the environmental footprint of agents on closed models, said Hausfather did good work but leaned on somewhat outdated figures. That is not surprising: academic research on the question is thin, and technology companies are reluctant to publish emissions numbers.
Hausfather's own reading is that on the scale of a personal life, energy use equivalent to keeping a few spare refrigerators permanently running is not catastrophic. But it is a new source of emissions arriving while global temperatures climb quickly and emissions targets slide further out of reach. It is also conspicuously larger than the fraction-of-an-almond figures Altman offers.
Hausfather notes that he uses AI and agent tools "more than most people," and that this may soon stop being true. Last week Meta introduced a personal AI agent that the company says is "created to work for the benefit of billions of people around the world." It is called Muse. Meta said each user will get "a separate computer in the cloud" that can keep working while the user is away, and that later this year it plans to connect Muse to its AI glasses. In the near future, someone using the glasses or Facebook may hand tasks to agents without noticing they have done it.
Here is where the almond starts to look like rhetoric rather than measurement. The unit was chosen because it produces a small answer; per-query comparisons work only while the query is the atom of usage, and the entire product roadmap is a bet that it will not be. The more interesting question is not what one prompt costs but whether agent demand has any natural boundary at all. Gamazaychikov's framing is the sharpest thing in this debate: every consumer technology before this one was capped by human attention, and agents are explicitly designed to run without it. Notably absent from every executive statement quoted here is any figure for what an agent session costs. Companies that can quantify almonds can quantify this, and have not.
The construction schedule shows what they expect. Compared with regular air travel or eating beef every day, one person's agent use still carries a small carbon footprint. But if Meta is planning for a world where everyone uses agents, the size of some of the data centers now going up makes sense. The Hyperion project in Louisiana will draw power from 10 gas power plants. The technologies trained in the data centers now proposed and under construction, Gamazaychikov says, will arrive in three to five years, and they will be a different kind of system, not a chatbot window.
The clean alternative exists on paper. Small nuclear plants could be a good source of carbon-free power for data centers, and several startups and developers picture a future in which data centers run alongside small modular reactors connected to the grid. The problem, as always with nuclear, is timing. No small modular reactor operates commercially in the United States, and despite decades of development only one model holds a license to sell. The Trump administration is trying to speed the sector up: the Department of Energy launched a pilot program for 11 startups that must hit a significant milestone this year, and at least a few have already cleared it. A long road to market remains. Data center developers who do not want to wait years for that proof are installing gas turbines now.
The local politics are moving faster than any of this. For Scientific American, Austyn Gaffney went to Memphis to report on protests against SpaceX's data centers. The Wall Street Journal has described states that granted data centers tax breaks and are now revisiting those deals. Texas station KERA News reports that data centers have pushed some Republican voters to think about leaving the party.
So the sequence is set: gas plants built this year, for systems that will not exist for three to five, to serve a demand curve nobody has measured, with the first serious accounting of agent energy use due out later this month — after the turbines have been ordered.