AI agents consume up to 137 times more power than chatbots
Analysis of AI agent power usage: confirmed 137 times higher consumption and latency than chatbots Also quickly assess the impact on data center costs by examining LLM calls and GPU idle time
A KAIST research team analyzed the power consumption and response times of AI agents and found that agentic AI uses far more electricity than conventional chatbots.
According to the study, while processing a single question, AI agents repeated more inference and tool calls on average, resulting in 9.2 times more LLM calls than the existing stepbystep reasoning method. As a result, response times were up to 154 times longer, and the amount of time GPUs remained idle while waiting for tasks also increased significantly.
In particular, based on a model with 70 billion parameters, AI agents used an average of 348 watthours (Wh) of electricity per question, which was 137 times higher than chatbots. The research team explained that these results could have an important impact on data centers' power demand and operating costs.