AI agents' hidden power appetite is fueling the data center boom
In her weekly Wired column Power Play, senior writer Molly Taft tackles a question she says readers keep asking: why are tech companies taking on billions of dollars in debt and building some of the world's biggest power plants just to run more data centers, when AI already looks capable enough? Her answer is that the industry has moved on from simple chatbot queries to agents: large language model based systems that, as her colleague Maxwell Zeff describes it, can generate hundreds of their own follow-up prompts to complete a task, running for hours and re-prompting themselves dozens of times to build out a website's pages, menus and datasets from a single instruction. That shift, she argues, is a major driver of the current buildout.
As an example of the scale involved, Taft cites OpenAI's announcement that a swarm of more than 10,000 agents, exchanging 2.7 million messages, had solved a longstanding math problem, a claim mathematicians have since pushed back on. Zeff estimated the processing power behind that swarm probably cost tens of millions of dollars in energy, though he says an exact figure is hard to pin down. The piece contrasts this with how AI companies usually talk about their environmental footprint: OpenAI CEO Sam Altman has claimed in a podcast interview that the water needed to grow one almond equals 38,000 ChatGPT queries, adding that people 'scarfing down 12 almonds at a time' do not think of that as a water problem, a calculation that has itself been disputed. Taft argues that framing individual queries this way obscures the much larger footprint of agents, whose energy use ranges from trivial tasks to a full day of autonomous coding by teams of parallel helper agents, with essentially no ceiling as tasks grow more complex.
Boris Gamazaychikov, co-founder and CEO of the research group Sustainable AI, tells Taft that unlike cars or Netflix streams, agent workloads are not capped by how many people are using them, which is exactly the decoupling AI leaders want: he cites talk of one-employee 'unicorns' backed by hundreds or thousands of agents working in the background. With companies disclosing little verifiable energy data, some outsiders are trying to estimate it themselves. Climate scientist Zeke Hausfather recently published a blog post calculating that his own heavily agent-based daily Claude usage may consume more energy than running two refrigerators, concluding that while this is not world-ending for his personal footprint, it is still a net new source of emissions at a time when global emissions targets are falling behind. Gamazaychikov, whose group plans to publish more precise agent-emissions research this month, told Taft that Hausfather's effort was reasonable but relied on somewhat outdated figures, reflecting how little independent academic work exists on the topic.
Taft points to Meta's newly launched personal AI agent, Muse, as a sign this usage could scale fast: the company says in a press release that Muse is 'built to work for billions of people worldwide' and will maintain a 'dedicated computer in the cloud' for each user even while they are offline, with plans to integrate it into Meta's AI glasses later this year. She connects this to the scale of projects like the Hyperion data center in Louisiana, which will be powered by 10 natural gas plants, and quotes Gamazaychikov's prediction that the technology being trained by data centers now under construction is three to five years away and 'a very different flavor than just the chatbot window.' In a reader mailbag section, Taft also addresses whether small modular nuclear reactors could power data centers: she says they could in principle, but none are yet operating commercially in the US, only one design has been licensed for sale despite decades of development, and while a Trump administration pilot program has pushed 11 startups toward a milestone this year, with at least a handful succeeding, most data center developers are not waiting for that industry to mature and are installing gas turbines now instead.
Key facts
- OpenAI said a swarm of more than 10,000 agents, sending 2.7 million messages, solved a longstanding math problem; mathematicians pushed back on the claim, and a colleague of the author estimates the processing power behind it probably cost tens of millions of dollars.
- Sam Altman has compared AI's water footprint to growing a single almond (38,000 ChatGPT queries), a framing the article says obscures the much larger and largely undisclosed energy cost of multi-step AI agents versus simple chatbot queries.
- Climate scientist Zeke Hausfather estimated his own agent-heavy daily Claude use may burn more energy than running two refrigerators; Sustainable AI's Boris Gamazaychikov called the effort reasonable but based on outdated figures.
- Meta's newly launched personal agent Muse is described in a company press release as built for billions of people and running a dedicated cloud computer per user, with AI-glasses integration planned later this year.
- The Hyperion data center project in Louisiana will run on 10 natural gas plants; separately, only one small modular nuclear reactor design has been licensed for sale in the US despite decades of development, so most developers are installing gas turbines now rather than waiting.
Why it matters
The piece reframes the data center boom: the industry's public messaging still leans on single-query comparisons like Altman's almond, but the actual driver of the buildout is agents, which can run unsupervised for hours and multiply a single request into hundreds of follow-up steps. That decoupling of compute demand from the number of human users, as Gamazaychikov puts it, is structurally different from past technology growth that was capped by how many people drove cars or streamed video, and it is what is pushing companies to build power plants alongside data centers rather than just server halls.
Who it affects
It affects anyone whose electricity bills, air quality or local politics intersect with new data center construction, since projects like the gas-fired Hyperion campus in Louisiana are sized for agent-scale compute rather than chatbot traffic. It also affects everyday consumers of Meta products: the article notes that people using Meta's glasses or Facebook may soon be delegating tasks to the Muse agent, described by the company as built for billions of users, without necessarily realizing how much background compute that involves. Researchers and journalists trying to estimate AI's environmental footprint are affected too, since they are working with self-reported, individual-query metrics from companies rather than agent-level data.
How to use it
The article's own reader question offers a concrete data point: small modular nuclear reactors are a plausible clean-power option for data centers in principle, but only one design is licensed for sale in the US and none operate commercially yet, so developers seeking power now are turning to gas turbines instead. For anyone trying to gauge AI's real energy impact, the practical takeaway is to treat single-query comparisons, like the almond-water figure, with skepticism, and to look for agent-level or task-level energy estimates, such as Hausfather's refrigerator comparison, as a more honest starting point even though those are still described as based on outdated data.
How solid is it
The column is built on named, on-record sources: Wired's own reporters, an outside sustainability researcher, and a climate scientist's public blog post, plus a company press release for the Muse claims. But several of the headline figures are explicitly approximate or contested: OpenAI's math-problem claim has been disputed by mathematicians, the cost of its agent swarm is a rough verbal estimate from a colleague rather than a measured figure, and Altman's almond calculation has been publicly disputed. Gamazaychikov also flags Hausfather's own energy estimate as based on somewhat outdated findings, which the article attributes to how little independent academic research exists on agent energy use.
Risks and caveats
The source does not give a name or date for the math problem OpenAI's agent swarm solved, an exact dollar figure for its energy cost, a date or venue for Altman's podcast remarks, or a numeric multiplier for how much more energy an agent task uses than a chatbot query. It also does not specify Muse's current user numbers, the exact DOE milestone the 11 small-modular-reactor startups were pursuing or how many met it, or further detail on the Hyperion project beyond its Louisiana location and 10 gas plants. Sustainable AI's promised more precise agent-emissions research had not been published at the time of writing, so the article's own energy estimates remain provisional.
“The people that are scarfing down 12 almonds at a time don't feel like they're doing something horrible from a water perspective for the most part.”
— Sam Altman, OpenAI CEO