Why AI Agents Guzzle So Much More Power Than a Simple Chatbot

Key points

• AI agents, software that carries out long multi-step tasks without human input, use far more computing power than a single chatbot query.
• OpenAI used a swarm of more than 10,000 agents sending 2.7 million messages to solve one maths problem, at a cost Maxwell Zeff of Wired AI estimates could be tens of millions of dollars in energy.
• Climate scientist Zeke Hausfather calculated that his own daily use of an AI agent may consume more energy than two refrigerators running continuously.
• Meta launched a personal AI agent called Muse, promising a dedicated cloud computer for each of its billions of users.
• Meta's Hyperion data-centre project in Louisiana will draw power from 10 natural gas plants.

An exasperated friend put it to me this way recently: "What on earth are they building all of these data centres for?" Fair question. Tech companies are taking on billions of dollars of debt and building some of the largest power facilities in the world, and the reason is not recipe searches.

Chatbots are yesterday's model. The real money is now in AI agents: software systems built on large language models (the technology behind tools like ChatGPT and Claude) that make their own decisions and carry out tasks over many steps, sometimes running for hours with no human involved.

What does an agent actually do?

Instead of answering one question, an agent breaks a job into hundreds of smaller steps and works through them alone. Ask it to build a website and it might spend hours writing pages, assembling datasets and rebuilding menus, re-prompting itself dozens of times. That is a very different workload from asking a chatbot what the weather is like.

The scale can get enormous. OpenAI recently announced that a swarm of more than 10,000 agents, sending 2.7 million messages between them, solved a long-standing maths problem. Mathematicians have since pushed back on how the result was framed. The energy cost of that single experiment was probably tens of millions of dollars, according to Maxwell Zeff, who covers AI infrastructure at Wired AI.

How much energy does personal agent use cost?

More than most people realise. AI companies rarely publish precise figures, which makes independent calculation hard. Climate scientist Zeke Hausfather did his own sums last month and estimated that his average daily Claude session, which leans heavily on agentic tasks, may use as much electricity as two refrigerators running at the same time. Boris Gamazaychikov, co-founder of Sustainable AI (a research and advisory group), says Hausfather's method was solid even if some underlying data was a little out of date.

For context, OpenAI CEO Sam Altman has suggested that a single ChatGPT query uses about as much water as one thirty-eighth of an almond. That framing covers a simple back-and-forth. An agent running for several hours is a completely different creature, and the almond maths does not stretch that far.

Meta launched its personal agent, Muse, promising a dedicated cloud computer for every user, running tasks even when your phone is in your pocket. The company plans to connect Muse to its AI glasses later this year. Millions of people could soon be running agents without noticing they have started.

"In other technological growth areas, we're constrained by how many people are driving a car or streaming Netflix," Gamazaychikov told Wired AI. "Now, this stuff is kind of decoupled from users." Tech companies are openly working toward a future where a business might have one human employee and hundreds of AI agents doing the rest. Realistic or not, that vision is what is driving the construction rush.

The carbon footprint of one person's agent use is still smaller than a beef-heavy diet or frequent flying. But it is a new source of emissions added on top of everything else, at a moment when global temperatures keep breaking records. The infrastructure being built today, including Meta's Hyperion gas-plant complex in Louisiana, is designed for a world of agents at scale. Our 2 September story on communities pushing back against AI data centres showed how quickly local opposition forms once the building starts; the agent energy story gives those communities sharper numbers to argue with.

What happens next?

Gamazaychikov puts it plainly: the technology that all this construction will train is still three to five years away. What comes out the other end will look nothing like a chatbot window.

The honest read here is that the industry's energy disclosures are still far too thin for anyone to make confident judgements about the real cost. Until companies publish agent-level figures rather than single-query snapshots, we are all doing arithmetic with a ruler someone has hidden most of.

Common questions

Is my personal use of ChatGPT or Claude a serious environmental issue?

For now, one person's chatbot habit is a small fraction of their overall carbon footprint. The concern is about agents running at scale across billions of users, which is what companies like Meta are now building toward.

Could nuclear power solve the data-centre energy problem?

Small modular reactors, compact nuclear plants that could sit next to a data centre, are a popular idea, but none operate commercially in the US yet and only one design has been licensed for sale. Most data-centre developers are not waiting: they are installing gas turbines now and hoping cleaner options arrive later.

Will I know if an app is using an AI agent on my behalf?

Not always. Meta's Muse is designed to run tasks in the background while your phone sits in your pocket. As agents get built into glasses and everyday apps, it will become harder to tell when one is working on your behalf.