The energy debate around AI is increasingly outgrowing the familiar comparison with a single chatbot query. WIRED argues that autonomous agents, which can break one request into hundreds of model calls and keep working for hours, offer a more useful explanation for the industry’s expanding data-centre and power plans.
There is no official definition of an agent in the article. Broadly, the term covers language-model systems designed to make decisions and execute a task with less continuous direction from a person.
One Request Can Become a Long Chain of Work
A user may ask an agent to build a website once, but the system can create dozens of prompts for itself while assembling pages, menus, features and datasets. A full-day coding job may also employ several parallel helper agents. The computing load therefore depends on the length and structure of the process, not simply the number of people typing questions.
One unusually large example came from OpenAI, which said more than 10,000 agents exchanged 2.7 million messages while working on a longstanding mathematics problem. Mathematicians disputed aspects of the company’s claim. WIRED’s Maxwell Zeff estimated that the effort probably consumed processing worth tens of millions of dollars, while emphasizing that an exact figure is difficult to establish.
That case is an outlier, not a representative consumer task. It nevertheless shows how quickly autonomous loops can multiply inference. The gap between a brief lookup and a coordinated agent swarm makes a single per-query energy figure a poor description of the emerging workload.
Energy Disclosure Has Not Kept Pace
Private AI companies disclose little consistent information about the environmental footprint of their products. Executives often describe individual queries, including Sam Altman’s claim that producing one almond uses as much water as 38,000 ChatGPT requests. WIRED notes that the calculation has been disputed.
Agents make such comparisons harder because workloads can range from short tasks to hours of autonomous activity. Boris Gamazaychikov, cofounder and chief executive of the research and advisory group Sustainable AI, says AI demand is becoming detached from the number of human users. A company imagined with one employee could still operate hundreds or thousands of agents in the background.
Climate scientist Zeke Hausfather attempted to estimate his own agent-heavy use. Drawing on several sources, he calculated that an average day with Claude might consume more electricity than two refrigerators. Gamazaychikov called it a strong effort but said the underlying findings were somewhat outdated, reflecting the limited academic research and weak corporate disclosure.
Hausfather does not present that personal total as catastrophic. He does describe it as a new source of emissions at a time when temperatures are rising and reduction goals are off track. The estimate is illustrative rather than a standardized measurement.
Mass Adoption Explains the Infrastructure Scale
Agent use may spread well beyond intensive early adopters. Meta says its Muse personal agent is intended for billions of people and will provide each user with a dedicated computer in the cloud that can continue working while the user is offline. The company plans to integrate Muse with its AI glasses later this year.
If ordinary social-media or glasses interactions start autonomous tasks in the background, users may generate agent workloads without thinking of them as separate AI sessions. Even a modest footprint per task becomes significant when multiplied across a very large population and repeated continuously.
WIRED connects that ambition with facilities such as Meta’s Hyperion data-centre project in Louisiana, which is expected to use ten natural-gas plants. Personal agent use may remain smaller than emissions from frequent flying or daily beef consumption, but the target scale helps explain why developers are building power infrastructure far beyond today’s chatbot traffic.
The article also notes that the technology intended for data centres now being proposed may be three to five years away. Reliable agent-level energy reporting is therefore missing precisely when long-lived generation and construction decisions are being made. The central uncertainty is no longer whether one answer uses much power, but how many unseen model calls each future request will set in motion.
