Source notes: This article is based on reporting from WIRED on agentic AI and data-center power demand, with additional industry context from the current public AI infrastructure race.

🤚 The Open-Palm Kilowatt

AI agents, those tireless software interns that do not sleep, unionize, or understand why procurement exists, are becoming the industry’s favorite explanation for the vast new data-center buildout. According to WIRED, the shift is not merely about answering chatbot questions faster. It is about systems that can break a user’s request into many smaller prompts, run for long stretches, call tools, generate code, revise plans, and generally behave like a junior consultant trapped inside a graphics card.

The old mental model was quaint: a person asks a chatbot for a pasta recipe, the model responds, and somewhere a server sighs elegantly. The new model is less “question and answer” and more “please assemble a digital task force and keep it busy until the cloud invoice develops a personality.” WIRED notes that agents can re-prompt themselves dozens of times while building something like a website, and that frontier labs increasingly treat agentic workflows as central to the next phase of AI.

One vivid example: OpenAI recently said a swarm of more than 10,000 agents exchanging 2.7 million messages helped solve a longstanding math problem, though mathematicians pushed back on the framing. Even with that caveat, the infrastructure point remains magnificently unromantic: millions of model calls require processing power. Processing power requires electricity. Electricity requires somebody, somewhere, to explain to a local community why the future sounds like a substation humming through dinner.

👐 The Two-Handed Refrigerator Invoice

The industry has often preferred to discuss AI’s footprint through adorable micro-comparisons: a single query, a drop of water, a fraction of a snack, perhaps a morally neutral almond wearing a tiny sustainability badge. But agentic AI makes those comparisons feel like judging a cruise ship by one cocktail napkin.

WIRED highlights the problem: agent tasks vary wildly. A simple request may be modest. A full-day autonomous coding session with parallel helper agents is not the same animal. Climate scientist Zeke Hausfather, attempting to estimate his own agent-heavy AI use, calculated that an average daily Claude session could consume more energy than powering two refrigerators. The exact math is uncertain, partly because the largest AI companies disclose environmental and operational metrics with the enthusiasm of a duchess revealing her offshore accounts.

This opacity matters because agents decouple compute demand from the number of human users. Streaming video has a practical relationship to people watching things. Cars have a relationship to people driving them. Agents, however, can multiply quietly in the background, each one asking the next little machine question like a committee of ghosts preparing a slide deck.

Meta has already described Muse, a personal AI agent built to work for billions of people worldwide, with dedicated cloud computing for users even when they are offline. That is not a chatbot; that is a private valet idling in a digital driveway. Multiply that by social networks, smart glasses, office suites, coding tools, customer-service platforms, and every enterprise product now required by law of venture capital to contain an “agent,” and the data-center gold rush begins to look less mysterious.

🌿 The Gentle Awakening

None of this means every AI agent is an environmental catastrophe in a waistcoat. Compared with regular flights, beef-heavy diets, and humanity’s long-standing hobby of setting hydrocarbons on fire for ambiance, personal agent use may still be relatively small. The issue is scale, opacity, and the luxury-grade confidence with which the sector builds first and explains later.

The companies are not constructing power-hungry campuses because someone needs a better limerick about Kubernetes. They are betting on a future where software does not merely answer but acts: books the trip, writes the app, monitors the supply chain, negotiates the calendar, drafts the lawsuit, summarizes the lawsuit, and then apologizes for hallucinating the judge.

That future may be useful. It may also be wildly inefficient. Both can be true, which is inconvenient for everyone who prefers their narratives pre-chewed and venture-backed.

👑 The Gold-Leaf Reckoning

The great comedy is that the AI industry sold the public on ethereal intelligence — weightless, frictionless, pure thought in the browser — while quietly becoming one of the most physical industries on Earth. Land, water, chips, transmission lines, gas turbines, tax breaks, nuclear fantasies, and local opposition now sit beneath the chatbot window like the basement kitchen under a five-star restaurant.

If agentic AI is the next computing platform, then its bill will not be paid only in subscription fees. It will be paid in infrastructure, grid stress, emissions accounting, and political fights over who gets cheap power and who gets the noise. The assistant may be virtual. The power plant is not.

“Your autonomous workflow has requested another autonomous workflow to optimize its autonomy.” — The Slap of Wisdom Department of Premium Load Forecasting, reviewing the champagne menu beside a backup generator