Research shows up to 136.5x electricity per query, slower performance, idle GPU usage, and potential 198.9 gigawatt data center demand.

A new peer-reviewed study by the Korea Advanced Institute of Science and Technology (KAIST) has argued that AI agents come with a far steeper energy bill than ordinary chatbots, and the gap is large enough to raise concerns about future data center demand and grid capacity.

Researchers presented their work at the 32nd IEEE International Symposium on High-Performance Computer Architecture, announcing findings that agent systems can consume up to 136.5 times more electricity per query, a difference they traced to the extra planning, tool use, and repeated model calls that autonomous agents perform.

By examining how AI agents use computing resources in real service-like conditions, the team have taken a quantitative look at the compute cost, latency, energy use, and broader power implications. In one test, an agent using a 70-billion-parameter language model averaged 348.41 watt-hours per query, compared with the far lower energy use of a conventional generative AI system doing simple question answering.

The study has also showed that agent workflows can be much slower than chatbot responses, with generation taking up to 153.7 times longer in some cases because the system keeps looping through decisions and external tools.

KAIST researchers say expensive graphics processing unit (GPU) hardware often sits idle for more than half the time while the agent waits on websites or apps to respond, which further wastes electricity.

The team subsequently modeled a world with 13.7bn AI agent requests per day, and estimated that data center power demand could climb to about 198.9 gigawatts. That is roughly half of average electricity consumption in the United States, and far above the scale of the multi-gigawatt AI data centers now being built.

Furthermore, the researchers believe that software tweaks alone will not solve the problem; the industry needs a broader redesign that includes AI models, chips, and data center power systems. As enterprises rush to deploy agentic AI for consumer and enterprise use, the study suggests the energy cost of that convenience may be much higher than many users realize.