Artificial intelligence (AI) agents consume up to 136.5 times more energy per query than conventional generative AI, a study has found.
KAIST said Thursday that a research team led by Chair Professor Yu Min-soo of the School of Electrical Engineering conducted the world's first systematic analysis of how much computing resources and power AI agents use in real service environments.
AI agents go beyond simply answering questions like large language models (LLMs) such as ChatGPT. They plan on their own and solve complex problems by using external tools such as internet search, calculators, and code execution. Their use is expanding across various fields, including software development, research support, and task automation, but the power and costs required to operate them as actual services have not been sufficiently understood.
The research team viewed AI agents as a new type of workload that data center servers and graphics processing units (GPUs) must continuously handle, and analyzed the amount of computation and energy consumption generated during their execution. As a result, AI agents were found to generate an average of 9.2 times more LLM calls than chain-of-thought (CoT), the existing step-by-step reasoning method.
Response time also increased significantly. AI agents' answer time rose up to 153.7 times. In particular, the time GPUs spent idle, unable to compute while external tools performed tasks, accounted for up to 54.5% of the total execution time. This inefficiency, in which expensive GPUs were not fully utilized, arose in the process of switching among multiple tools to solve complex problems.
Energy consumption was far greater than that of conventional generative AI. According to the team's analysis, an AI agent using a large language model with 70 billion parameters consumed an average of 348.41 Wh of power to process a single query. This is 136.5 times higher than generative AI using the typical single question-and-answer method.
The team expanded the analysis to the data center scale. Assuming a future environment in which 13.7 billion AI agent requests occur per day, data center power demand was estimated to reach about 198.9 GW. This far exceeds the multi-gigawatt-scale AI data centers currently being pursued by countries around the world, and corresponds to about half of the average total power consumption of the United States.
This study shows that AI competitiveness is expanding beyond model performance to the efficiency of AI semiconductors, data centers, and power infrastructure. Going forward, co-design that optimizes AI models, semiconductors, data centers, and power grids together—rather than merely building smarter AI models—is expected to become important.
"This study is the first case to quantitatively present how much power and cost are needed to implement and maintain AI intelligence, beyond AI becoming smarter," Chair Professor Yu Min-soo of KAIST said. "In an era when AI agents become widespread, an approach that integrates and optimizes not only AI data centers but also AI agent models and power infrastructure will become important."
The study was carried out with Kim Ji-in, a doctoral student in KAIST's School of Electrical Engineering, as first author. The research results were presented in February at IEEE HPCA, an international conference in the field of computer system design. The team made the AI agent implementation technology and benchmark used in the paper available as open source so it can be used in follow-up research.
