AI agents, not chatbots, drive data center buildout: Wired

The article traces a shift in how AI companies and their critics think about energy use. For years, the standard defense was the single chatbot query: OpenAI CEO Sam Altman compared the water needed for one almond to 38,000 ChatGPT queries, a framing that made individual AI use seem trivial. That comparison is now being challenged from two directions. First, agents change the math—Wired's Maxwell Zeff describes agents that might run for hours and re-prompt themselves dozens of times to build a website, and an OpenAI experiment with more than 10,000 agents sending 2.7 million messages to solve a math problem burned through a lot of processing power. Second, the lack of disclosure from private AI companies means outsiders are doing their own estimates, like climate scientist Zeke Hausfather, whose blog post put his daily Claude use at more than two refrigerators' worth of energy.

The article connects that personal-scale estimate to the industrial-scale buildout. Meta's new Muse agent is described as built to work for billions of people, each with a dedicated cloud computer, and Meta plans to integrate it with its AI glasses later this year. That vision helps explain projects like the Hyperion data center in Louisiana, which will be powered by 10 natural gas plants. As Boris Gamazaychikov of Sustainable AI puts it, the technology being trained by data centers proposed and built right now is three to five years away and will be a very different flavor than the chatbot window.

What the outcome hinges on is whether agent use becomes as widespread as Meta envisions—and whether anyone can measure its footprint. The article notes a major dearth of information around agent energy use, and Gamazaychikov's group plans to release research later this month with more precise calculations for agents running on closed models. Data center developers, meanwhile, are not waiting for small modular reactors to mature; they're installing gas turbines now, which suggests the near-term emissions path is already being locked in.