AI Agents Are Thirsty for Power
wired ·
The article says Silicon Valley's data center construction is being driven by a shift from ordinary chatbot queries to resource-intensive AI agents—roughly LLM-based systems that autonomously decide how to execute tasks. Unlike a simple question-and-answer, an agent can give itself hundreds of small prompts based on a user's original request; for example, building a website might take hours and require dozens of self-prompts to create pages, menus, and datasets. Agents are now central to frontier labs' work: OpenAI announced that a swarm of over 10,000 agents sending 2.7 million messages solved a longstanding math problem, though mathematicians pushed back and the messages consumed enormous processing power. The article notes private AI companies disclose little about environmental metrics; Sam Altman's podcast claim that one almond's water equals 38,000 ChatGPT queries has been disputed, and agents are far more energy-intensive with little reliable data. Climate scientist Zeke Hausfather estimated his daily Claude session, which leans heavily on agents, might use more energy than powering two refrigerators, though Sustainable AI CEO Boris Gamazaychikov said the math was based on somewhat outdated findings. Hausfather concluded his AI use was not world-ending personally but represented net new emissions while global temperatures rise and emissions reduction goals are off track, and it is much larger than Altman's almond-sized metric. Meta's new personal agent Muse, described as built for billions of people with a dedicated cloud computer per user that works offline, shows how agent use could spread, especially with planned AI glasses integration later this year. The piece ties this demand to huge data centers such as Louisiana's Hyperion project, powered by 10 natural gas plants, and notes that small modular nuclear reactors could someday provide carbon-free power but none operate commercially in the US, so developers are installing gas turbines now.
wired