AI Agents Use Up to 136.5 Times More Energy as GPUs Sit Idle

Unlike conventional question-and-answer AI systems that generate a response in a single pass, AI agents repeatedly call large language models, plan steps and use external tools such as search. Research from the Korea Advanced Institute of Science and Technology (KAIST) shows that GPUs struggle to remain productive while waiting for those tools to respond, making energy efficiency a major risk for large-scale agent deployment and data-center power planning.

The latest measurements found that a single AI-agent query can consume as much as 136.5 times the energy used by conventional question-and-answer AI, while GPUs sit idle for more than half of the processing time. Available information did not specify the study's publication date, the models tested or electricity costs. Still, the data highlight the energy burden created by tool calls and waiting times, suggesting widespread adoption could place further strain on global power infrastructure.