Complex AI agents that independently plan and execute tasks consume up to 136.5 times more electricity per query than simple question-and-answer chatbots, according to new research. The quantitative analysis suggests that competitiveness in the artificial intelligence era will extend beyond model performance to encompass the efficiency of data centers and power infrastructure.
The Korea Advanced Institute of Science and Technology (KAIST) announced on the 5th that a research team led by Professor Yoo Min-soo of the School of Electrical Engineering had analyzed the computational resource usage and energy consumption of AI agents under simulated real-world service conditions. The team defined AI agents not as simple programs but as a new type of workload that data center servers and GPUs must continuously process, and comprehensively examined the performance degradation factors and power consumption arising in the process.
The analysis showed that, unlike conventional generative AI, AI agents do not stop at generating a single answer. They formulate plans to solve problems, request additional information, and iteratively incorporate results from external tools such as search or code execution. This repeated invocation of large language models (LLMs) caused total response times to increase by up to 153.7 times compared to conventional methods.
GPU utilization efficiency also dropped sharply. While the AI was using external tools for tasks like search or computation, GPUs remained idle with no computation to perform. According to the research team, GPU idle time accounted for up to 54.5% of total execution time. This means that even when deploying expensive AI semiconductors, substantial idle time occurs during actual processing.
The difference in power consumption was even more pronounced. Based on a large language model with 70 billion parameters—comparable to current commercial AI services—an AI agent consumed an average of 348.41 watt-hours (Wh) of electricity to process a single query. That is up to 136.5 times more than a standard generative AI Q&A interaction.
The research team also ran simulations of a future environment where AI agent usage becomes ubiquitous. Assuming a scenario with 13.7 billion AI agent requests per day, the required data center power demand was estimated at approximately 198.9 gigawatts (GW). “This exceeds the multi-gigawatt-scale AI data centers currently being pursued by countries around the world,” the team noted, “and represents roughly half of the total average electricity consumption of the United States.”
The study demonstrates that the AI industry’s focus is shifting from “smarter AI” to “more efficient AI.” The researchers forecast that co-design—jointly optimizing not only AI model performance but also AI semiconductors, servers, cooling systems, and power grids—will become an essential strategy going forward.
“This study goes beyond simply showing that AI is getting smarter; it quantitatively demonstrates how much power and cost are required to implement and sustain that intelligence,” emphasized Professor Yoo Min-soo. “In an era where AI agents become commonplace, an integrated co-design approach that optimizes not only AI data center infrastructure but also AI agent models and power infrastructure together will become increasingly critical.”