Tech companies are taking on billions of dollars in debt and building massive power plants to construct new data centers. People often wonder why these facilities are necessary if artificial intelligence already makes breakthroughs. Simple chatbot queries are an increasingly outdated way of thinking about how artificial intelligence works. The current focus centers on agents. While there is no official definition, agents are large language model systems designed to make autonomous decisions to execute tasks.
Maxwell Zeff, who writes the weekly Model Behavior newsletter, explains the shift. Rather than asking a simple question and getting an answer, agents give themselves hundreds of small prompts based on an original user question. Zeff notes that if someone asks an agent to build a website, it might run for hours. It reprompts itself dozens of times to build features, web pages, menus, and datasets.
Agents sit at the heart of frontier labs and perform extraordinary tasks. OpenAI recently announced that a swarm of over 10,000 agents sending 2.7 million messages solved a longstanding math problem. Mathematicians pushed back on these claims. Even so, this process burned through a massive amount of processing power and energy. The cost reached tens of millions of dollars, though exact figures are difficult to determine.
Private artificial intelligence companies historically withhold environmental metrics. CEOs often point to single queries as a measure of resource use. OpenAI CEO Sam Altman claimed in a podcast interview that the water needed to harvest a single almond equals 38,000 ChatGPT queries. This calculation has faced disputes. Altman remarked that people eating a dozen almonds do not feel they are harming the water supply.
Introducing agents makes these calculations much more complex. There is a severe lack of information regarding agent energy use. Tasks range from simple jobs to a full day of autonomous coding involving parallel helper agents. Boris Gamazaychikov, co-founder and CEO of Sustainable AI, notes that traditional technology growth areas are constrained by how many people drive cars or stream Netflix. He points out that agent technology is decoupled from users, matching what artificial intelligence leaders want.
Leaders discuss unicorns featuring a single human employee supported by hundreds or thousands of agents working in the background. This vision explains the massive rush to construct data centers. Without reliable data from companies, enthusiasts attempt their own math. Climate scientist Zeke Hausfather authored a blog post last month calculating his personal artificial intelligence use, which relies heavily on agents.
Hausfather calculated that his average daily Claude session consumes more energy than two refrigerators. Gamazaychikov noted that Hausfather made a good effort, though his math relied on slightly outdated findings. This reflects the lack of academic work and the opaqueness of tech companies regarding emissions metrics. Hausfather concludes that his usage is not world-ending on a personal scale, but represents a net new source of emissions during a climate crisis.
Meta rolled out a personal artificial intelligence agent built to work for billions of people worldwide. Named Muse, the company stated it maintains a dedicated cloud computer for each user that operates even when offline. Meta plans to integrate Muse with artificial intelligence glasses later this year. Users interacting with glasses or Facebook might soon outsource tasks to agents without realizing it.
The carbon footprint for personal agentic use remains relatively small compared to taking regular flights or eating beef daily. If Meta envisions everyone using an agent, it explains massive data center projects like the Hyperion project in Louisiana, which will run on ten natural gas plants. Gamazaychikov states that the technology trained by proposed data centers is three to five years away and will look different from a chatbot window.
Regarding power sources, small nuclear reactors could provide carbon-free energy for data centers. Startups and developers envision data centers coupled with small modular reactors running off the electric grid. The issue remains the timeline. No small modular reactors operate commercially in the United States, and only one model has been licensed for sale. The Trump administration is helping the industry mature, including a Department of Energy pilot project for 11 startups to hit milestones this year.
At least a handful of startups reached that milestone before starting the journey to market. Data center developers often prefer not to wait years and are installing gas turbines now instead of waiting for a nuclear utopia. Austyn Gaffney reported for Scientific American on the backlash against SpaceX data centers in Memphis. The Wall Street Journal reported on states walking back tax breaks for data centers. KERA News covered how data centers are making Republican voters in Texas consider leaving the party.
Developers will continue installing immediate power solutions while long-term energy technologies remain in development over the next few years.
