AI Coding Bots Now Eat 43% of Supercomputer Time

Supercomputers are the giant, building-sized machines that scientists use to simulate weather, design drugs, and train AI. Until now, they were built around human habits: a researcher logs in, runs a job, waits, checks results, goes home. A new paper reports that AI coding agents (AI that can take actions on a computer by itself) have quietly become real users of these machines — and they don't behave like people at all.

In the researchers' measurements, people running coding agents made up just 19.5% of users, but accounted for 55.8% of job submissions, 29.1% of processor time, and 42.7% of graphics-chip time. Agents fire off commands 20.8 times faster than humans, break work into tiny "try, change, run again" loops, and chase open-ended goals by trial and error — straight through nights and weekends. That clogs the scheduling system, hammers the file storage with endless small reads and writes, and wastes effort re-learning things earlier sessions already figured out.

There are safety worries too. The paper flags "prompt injection" — sneaky hidden instructions that can trick an AI into doing something its owner never intended — as a new risk on shared research machines, along with the difficulty of enforcing who is allowed to do what. Today's systems weren't designed to check whether the thing typing is a person or a program.

The authors argue against both obvious fixes: banning agents, or pretending they're just ordinary users. Instead, they call for "co-design" — rebuilding supercomputers so AI agents are first-class citizens that know the rules of the facility, share what they've learned, and can be held accountable. Expect this debate to shape how much computing power, money, and electricity the AI boom really consumes.