Researchers at the Korea Advanced Institute of Science and Technology (KAIST) have estimated the "hidden costs" of agent AI, and found that such systems can consume up to 136.5 times more energy per request than conventional generative models. Reason - in the logic of work: agent constantly refers to basic model, generating new answers at each step of reasoning, which creates multiplier effect. On average, an AI agent request consumes 348.41 Wh - the equivalent of a LED lamp for 24 hours.

In addition to energy consumption, agents have a significant response delay: task completion can take up to 153.7 times longer than the standard request. At the same time, GPUs sit idle for up to 54.5% of the time while agents coordinate their actions, creating a level of inefficiency that is not present in simpler AI scenarios.

The authors of the study modeled a scenario in which AI agents handle 13.7 billion requests per day (comparable to Google search). Without growth in energy efficiency, this would create demand for ~198.9 GW of capacity - about half of current US energy consumption.