AI Agents Are Thirsty for Power

|

,

In a recent Power Play column for WIRED, senior writer Molly Taft explains why tech companies keep building so many data centers. Her answer is that the industry is moving from simple chatbot questions to AI agents, which use far more computing power, and that shift is helping drive the build-out.

Key highlights

  • Agents, not chatbots. Taft notes that there is no official definition, but agents are roughly systems built on large language models that make their own decisions to carry out a task. Her colleague Maxwell Zeff, who writes WIRED’s Model Behavior newsletter, explains that a single request can lead an agent to prompt itself hundreds of times. For example, an agent asked to build a website might run for hours.
  • Big jobs, big energy costs. Taft cites OpenAI’s recent announcement that a swarm of more than 10,000 agents, sending 2.7 million messages, solved a longstanding math problem. Mathematicians pushed back on the claim. Zeff estimates the energy involved was probably worth tens of millions of dollars, though the exact figure is hard to pin down.
  • Per-query numbers miss the picture. AI companies rarely share environmental metrics, and CEOs often rely on single queries for comparisons. Taft cites Sam Altman’s podcast claim that growing one almond requires 38,000 ChatGPT queries, a calculation that others have contested. The complexity of agent tasks, ranging from simple jobs to a full day of coding with parallel helper agents, makes such comparisons challenging.
  • Demand is no longer tied to users. Boris Gamazaychikov, co-founder and CEO of Sustainable AI Group, tells WIRED that AI use is becoming decoupled from the number of people using it. Growth in cars or streaming depends on how many people use them. AI leaders instead describe companies with a single employee and many agents working in the background.
  • One heavy user’s math. Climate scientist Zeke Hausfather estimated that his daily Claude session consumes more energy than running two refrigerators. While he acknowledges that this isn’t a significant amount in his personal life, he emphasizes that it still contributes to “net new emissions” at a time when emissions targets are falling short. Gamazaychikov commends Hausfather’s efforts but points out that he relied on somewhat outdated research. His team plans to release more accurate estimates later this month.
  • Agents for billions of people. Meta recently introduced Muse, a personal agent that the company claims will allocate a dedicated cloud computer for each user and continue functioning even when offline. Taft highlights that the use of personal agents still has a smaller carbon footprint compared to frequent flying or consuming beef daily. Nevertheless, the prospect of widespread agent usage provides an explanation for projects like Meta’s Hyperion data center in Louisiana, which will be powered by 10 natural gas plants.

Answering a reader, Taft writes that small modular reactors could supply carbon-free power for data centers. None are operating commercially in the US yet, though, so many developers are installing gas turbines now instead of waiting.

Acknowledgement: This summary was generated using AI writing tools as part of my AI exploration process. For more information, check out the original articles listed below.

Related Resource Links