It started with an exasperated friend: “What on earth are they building all of these data centers for?” It’s a question I’ve heard a lot lately, and I understand why. There’s a strange disconnect happening right now. On one hand, we keep hearing that artificial intelligence is already transforming medicine, science, education, and just about every other field you can name. On the other hand, tech companies are taking on billions of dollars in debt and building some of the largest power plants the world has ever seen, all to fuel more and more data centers. It feels excessive, confusing, and, for a lot of people, a little frightening. You might think the answer is that these companies need more computing power to make chatbots better at answering your questions about banana bread recipes or weekend road trips. But that’s an increasingly outdated way to understand what’s actually happening inside this industry. The simple chatbot era is fading. The new wave is all about “agents,” and that shift is a big part of what’s driving this seemingly crazy buildout. If you’ve ever felt like the data center boom doesn’t quite make sense, you’re not alone. But once you understand what agents are and what they’re being asked to do, the picture starts to become clearer—and arguably even more unsettling.
So what exactly is an AI agent? There’s no universally agreed-upon definition, but here’s a rough way to think about it: an agent is a large language model–based system designed to make autonomous decisions in order to complete a task, rather than simply responding to a single prompt and stopping. In other words, instead of asking a chatbot a question and getting an answer, you give an agent a goal, and it goes off to figure out how to achieve that goal on its own. My colleague Maxwell Zeff, who writes a weekly newsletter about this world, explains it beautifully: “Rather than asking an AI chatbot a simple question and answer, these agents can give themselves hundreds of small prompts based on a user’s original question.” He gives a concrete example: if someone asks an AI agent to build them a website, the agent might run for hours, prompting itself dozens of times to build different web pages, menus, and datasets that power the thing. It’s not just answering a question. It’s doing something closer to actual work—planning, stepping back, reassessing, and executing a series of tasks that all add up to a finished product. That’s a fundamentally different way of using AI, and it demands a fundamentally different amount of computing power. Instead of a few split-second calculations, you’re looking at thousands or even millions of computations happening over an extended period. And each of those computations requires energy. Lots of it. This is the silent engine behind the data center boom, and it’s changing the way tech companies think about infrastructure, growth, and the environment.
The frontier AI labs are now pouring enormous resources into agents, and some of the early results are both astonishing and terrifying. Recently, OpenAI announced that a swarm of more than 10,000 agents sending 2.7 million messages had solved a longstanding math problem. Now, let’s pause there. Mathematicians pushed back on the claim, arguing that the “solution” wasn’t quite what it was made out to be. But even setting aside the specifics of the math problem, the sheer scale of what happened is worth sitting with: ten thousand autonomous systems, exchanging millions of messages, working together toward a single goal. That is not your chatbot. That’s a digital workforce, and it came with a serious price tag in terms of raw processing power. My colleague Max estimates that this kind of operation probably consumed tens of millions of dollars worth of energy—no one can say exactly how much, because the companies involved aren’t releasing precise figures. But think about that for a moment. Tens of millions of dollars in electricity for a single, relatively academic exercise. And this wasn’t an anomaly. AI labs are deeply committed to pushing past known boundaries, even if it means throwing unusual amounts of computing power at problems that are theoretically unsolvable. They want to prove that AI can do things human intelligence alone cannot, and they’’re willing to burn through astonishing resources to make that point. It’s an exciting vision, but it’s also one with a growing environmental footprint that’s barely being discussed in public discourse.
One of the frustrating things about this moment is how little transparency we’’vve gotten from the companies at the center of it all. Historically, private AI companies have been very choosy about what they disclose when it comes to environmental metrics. They might talk about hypothetical single queries, because that makes the resource use sound tiny, but they rarely talk about the full picture: the training runs, the millions of daily active users, the dozens of new features being developed behind closed doors, or the massive agentic systems that are now humming along in data centers somewhere in the middle of nowhere. OpenAI CEO Sam Altman recently leaned into a very specific comparison on a podcast: he claimed that the water use needed to harvest a single almond amounted to 38,000 ChatGPT queries. The implication was that you shouldn’’t worry too much about asking ChatGPT something, because it’s way less environmentally costly than eating a nut. Nevermind that the calculation has been widely disputed, and nevermind that Altman is, of course, deeply invested in making people feel good about using his company’’’s products. There’s something almost absurd about the comparison, because it reduces AI’s environmental impact to a tiny everyday action while completely ignoring the massive industrial-scale energy use that’s happening in the background. “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,” he said, as if to say: don’’’t worry, this isn’’’t on you, and it shouldn’’’t be a big deal. But the arrival of AI agents blows this entire rhetorical strategy apart.
When you introduce agents into the equation, the energy calculus changes dramatically. A simple query might take a fraction of a penny’’’s worth of electricity, depending on who you ask. But an agent that’s “working” for an hour—or a day—or a week—can consume thousands of times more energy. Some agents are doing simple tasks that wrap up in a few seconds. Others are essentially small teams of parallel “helper” agents, all working together on a complex project, prompting each other, spinning off sub-tasks, re-prompting themselves and each other in loops that can run for hours or days. There’s a massive gulf in power use between these different applications, and no clear way for an outside observer to measure any of it. Even energy experts are struggling to track what’’’s happening, because the industry is moving so fast and guarding its data so carefully. Boris Gamazaychikov, the co-founder and CEO of Sustainable AI, a research and advisory group, puts it in stark terms: “In other technological growth areas, we’’rere constrained by how many people are driving a car or streaming Netflix. Now, this stuff is kind of decoupled from users—and if you listen to AI leaders, I think that’’’s what they want. They’’’re talking about unicorns that have one employee.” That’s a chilling idea. In the old model, energy use was tied to human behavior. More people driving, more energy burned. More people streaming, more energy burned. But in the new model, a tiny team of humans could orchestrate a massive swarm of agents that run 24/7, using energy constantly, without any individual person making a choice to “drive” or “play something” in that moment. It’s automated, ambient, and constantly expanding—and nothing is putting the brakes on it.
So what does this all mean for the rest of us? For one thing, it means the data center buildout isn’t going to slow down anytime soon. It means energy demand from AI is going to keep growing, potentially without limits, as agents become more capable and more independent. It means communities across the country are going to keep feeling the strain of data centers on their water supplies, their local grids, and their electric bills, while the companies building those centers continue to dodge tough questions about accountability and environmental impact. It also means we, as ordinary people, need to start asking better questions. Instead of asking “What are they building all these data centers for?”, maybe we should be asking: “Who is going to be accountable for the energy that those data centers use?” “What happens when AI agents are running millions of tasks every minute, with almost no human oversight?” “Why are companies so reluctant to tell us the real numbers?” The promise of AI agents is genuinely exciting: they could discover new medicines, design sustainable technologies, solve intractable problems that have puzzled human minds for centuries. But the way they’rere being built right now runs the risk of creating enormous ecological damage that we’’’ll all have to live with, while the benefits flow mostly to a handful of enormous companies. That doesn’’’t have to be the story. With better transparency, smarter regulation, and a more open conversation about energy use, it’s possible to build a future where AI’s advantages are measured not just in breakthroughs and stock value, but in whether they actually make our lives and our planet better. But that future depends on ordinary people asking the hard questions—and refusing to accept a single almond as a satisfying answer.
