From WIRED · Molly Taft · · 3m
Why AI Agents Eat So Much Power
Chatbots were cheap. Agents that prompt themselves for hours are what the new gas plants are actually for.
In brief
Chatbots were cheap. Agents that prompt themselves for hours are what the new gas plants are actually for. Data center buildouts aren't sized for your chatbot questions. They're sized for a future where every person runs agents that work all day without them. Originally reported by WIRED.
What are they building
Readers keep asking why tech companies are taking on billions in debt for data centers and some of the world's biggest power plants. The answer is not recipe search.
“What on earth are they building all of these data centers for?”
Chatbots are the old model
Thinking about AI as a question in, an answer out is already outdated. The buildout is aimed at agents, not at people typing queries into a box.
Agents prompt themselves
An agent is an LLM system that makes autonomous decisions to finish a task. Ask it for a website and it may run for hours, re-prompting itself dozens of times.
“these agents can give themselves hundreds of small prompts based on a user's original question”
Ten thousand agents, one problem
OpenAI said a swarm of 10,000 agents traded 2.7 million messages to crack a longstanding math problem. Mathematicians disputed the claim. The compute bill was real either way.
Tens of millions in energy
Nobody outside the lab can price that math run exactly, but the estimate lands in tens of millions of dollars of energy. Companies stay choosy about what they disclose.
The almond defense
Sam Altman compared one almond's water use to 38,000 ChatGPT queries, a calculation others have disputed. It frames AI's footprint as something snack-sized.
“The people that are scarfing down 12 almonds at a time don't feel like they're doing something horrible from a water perspective”
Single queries miss the point
Agents run anywhere from a quick job to a full day of autonomous coding with parallel helper agents. That gulf is enormous, and almost nobody has measured it.
Demand cut loose from users
Netflix and driving scale with the number of humans doing them. Agent compute does not. AI leaders openly talk about one-employee companies.
“Now, this stuff is kind of decoupled from users, and if you listen to AI leaders, I think that's what they want.”
Two refrigerators a day
Climate scientist Zeke Hausfather ran the numbers on his own agent-heavy use and got more power than two fridges per daily Claude session. His sources were somewhat dated.
Small, but net new
Hausfather calls a few spare fridges survivable in his own life. His worry is that it is a fresh source of emissions while reduction targets already slip.
“also represents a net new source of emissions, at a time when global temperatures are skyrocketing”
An agent for everyone
Meta's new agent, Muse, keeps a dedicated cloud computer running per user even when that user is offline, with AI glasses integration coming later this year.
“built to work for billions of people worldwide”
Gas now, nuclear later
Meta's Hyperion site in Louisiana runs on 10 natural gas plants. Small modular reactors could power data centers cleanly, but none operate commercially in the US yet.
“The technology that's going to be trained by the data centers that are being proposed and built right now is three to five years away”
What it means
Data center buildouts aren't sized for your chatbot questions. They're sized for a future where every person runs agents that work all day without them.
