KAIST announced that the team led by Distinguished Professor Minsoo Yoo of the School of Electrical Engineering has, for the first time in the world, quantitatively analyzed how much computational resources and power AI agents use under real service conditions.
The research results were presented last February at the 32nd IEEE HPCA, an international conference in computer system architecture.
An AI agent based on a large language model with 70 billion parameters used an average of 348.41 watt-hours (Wh) to process a single question — up to 136.5 times higher than the simple question-and-answer mode of conventional generative AI.
Response times increased up to 153.7 times longer. GPUs sat idle for as much as 54.5% of total execution time while waiting for external tool responses.
Assuming a future environment with 13.7 billion agent requests per day, data center power demand could reach approximately 198.9 gigawatts — about half of the United States’ average total power consumption.
The paper title is “The Cost of Dynamic Reasoning: Demystifying AI Agents and Test-Time Scaling from an AI Infrastructure Perspective” by Jiin Kim, Byeongjun Shin, Jinha Chung, and Minsoo Rhu.
doi.org/10.1109/HPCA68181.2026.11408569
The research team released benchmark code open source.