KAIST study: AI agents use 136.5 times more power than chatbots

A first-of-its-kind study from South Korea's Korea Advanced Institute of Science and Technology has quantified the energy cost of AI agents, finding that these autonomous systems consume up to 136.5 times more electricity per query than conventional generative AI chatbots — a gap that researchers warn could strain global power grids as agent-based AI goes mainstream.

The Energy Gap

The study, led by KAIST professor Rhu Min-soo and published over the weekend, measured the power draw of AI agents — systems that go beyond answering a single question to independently plan and execute multi-step tasks like managing a budget or booking travel. Unlike a standard chatbot, which processes a query and stops, an AI agent repeatedly loops through its underlying language model, searching the web, running calculations, and issuing commands on its own.koreatimes+1

That autonomy carries a steep cost. Using a 70-billion-parameter large language model — comparable to current commercial AI services — the researchers found that a single complex agent query consumed an average of 348.41 watt-hours of electricity. In benchmark testing, the agent framework known as Reflexion on Meta's Llama-3.1-Instruct 70B consumed 136.5 times more GPU energy per query than a standard single-turn chatbot baseline, while another framework called LATS used 62.1 times more.letsdatascience+2

The researchers also found that agent queries take up to 153.7 times longer to generate than chatbot responses, and that expensive GPU chips sit idle more than half the time, burning electricity while waiting for external tools and websites to respond.koreatimes

A Grid Under Pressure

The study's most alarming projection concerns scale. If AI agent usage grows to an estimated 13.7 billion daily requests worldwide, total data center power demand could reach roughly 199 gigawatts — approximately half of the average electricity consumption of the entire United States.aintelligencehub+1

That figure dwarfs current data center loads. U.S. data centers consumed about 176 terawatt-hours in 2023, roughly 4.4 percent of national electricity use, according to the Lawrence Berkeley National Laboratory. The International Energy Agency projects global data center electricity consumption will more than double to around 945 terawatt-hours by 2030, but the KAIST study suggests agentic AI workloads could accelerate demand far beyond those forecasts.iea+1

A Call for Redesign

Rhu warned that the tech industry cannot rely solely on smarter software to address the problem. Instead, he called for a ground-up redesign of AI models, microchips, and data center power infrastructure to handle the workload that agentic AI will impose. The study arrives as major technology companies race to deploy AI agents across enterprise and consumer applications, raising fresh questions about whether the energy grid can keep pace with AI ambition.koreatimes