Artificial intelligence (AI) is rapidly evolving beyond simply answering questions into the era of “AI agents,” which can perform complex tasks by autonomously carrying out multi-step reasoning and using external tools. However, the extent to which these advances require additional resources and power has not been properly quantified until now.
A research team led by KAIST School of Electrical Engineering Chair Professor Minsoo Rhu has announced the first quantitative analysis of AI agents’ computational cost, response latency, energy consumption, and their broader impact on data center power demand. The findings were presented at IEEE HPCA 2026, one of the most prestigious conferences in the field of computer architecture, drawing significant attention from the academic community.
The research team focused on the fact that AI agents do not merely increase computational workload, but also impose a fundamentally new burden on data center infrastructure. Unlike conventional “chain-of-thought” reasoning, which proceeds step by step in a manner similar to human reasoning, AI agents operate by repeatedly invoking large language models (LLMs) throughout the task execution process.
The analysis found that AI agents triggered, on average, 9.2 times more LLM invocations than conventional approaches, while processing latency increased by as much as 153.7 times. In addition, during tool-use phases, GPUs inevitably remained idle, with idle time accounting for up to 54.5% of the total execution time. This means that as tasks become more complex, the inefficiency of underutilizing expensive GPUs becomes increasingly severe.
When extrapolated to the scale of data centers, the gap in power consumption becomes even more pronounced. An AI agent using a 70-billion-parameter LLM, comparable to the scale of today’s commercial AI services, consumed an average of 348.41 Wh to process a single query. This is approximately 136.6 times higher than the 2.55 Wh consumed by a conventional one-shot question-answering approach.
The resulting power burden on data centers could grow exponentially. If at least 71.4 million daily active ChatGPT users were to shift from conventional question-answering to AI agents, the required power would jump from 7.6 MW to 1 GW. If approximately 13.7 billion Google searches per day were to be handled by AI agents in the future, the required power would amount to 198.9 GW per day. This is a massive figure equivalent to nearly half of the average total power load of the United States, which stands at 476.9 GW.
“The significance of this study lies in showing that improving AI agent capabilities does not merely require more computation, but also creates a new class of burden across data center infrastructure,” the research team said. “Future research on AI agents must consider not only how to build smarter agents, but also how to operate them efficiently within limited infrastructure and power budgets.”
Kiyoung Choi, former professor at Seoul National University’s Department of Electrical and Computer Engineering and former Minister of Science and ICT, noted that in addition to the challenges surrounding AI data centers, semiconductor shortages, and material constraints, power supply is also a highly difficult issue. He emphasized the significance of this paper in sounding an alarm that endlessly expanding data center scale and cost in response to AI agents’ demands may not be sustainable.
The study was conducted by Master’s student Byungjun Shin, MS-PhD integrated student Jinha Jung, and PhD student Jiin Kim.
※ Paper title: The Cost of Dynamic Reasoning: Demystifying AI Agents and Test-Time Scaling from an AI Infrastructure Perspective
※ GitHub technical open source: https://github.com/VIA-Research/AgentBench
This work was supported by the SW Computing Industry Core Technology Development Program, including the SW Starlab Program, funded by the Ministry of Science and ICT and the Institute of Information & Communications Technology Planning & Evaluation, as well as by the Samsung Science & Technology Foundation.
Major Media Coverage
- 이데일리: “AI 에이전트, 질문 한 번에 전력 136배 더 쓴다”…KAIST, 세계 첫 정량 분석 (2026.7.5)
- 헤럴드경제: “챗GPT보다 전력소모 136.5배↑”…KAIST, AI 에이전트 ‘숨은 전력 비용’ 첫 규명 (2026.7.5)
- 파이낸셜뉴스: 똑똑하다는 ‘AI 비서’, 알고 보니 전기먹는 하마? 전력 소모량 136배 많아 (2026.7.5)
- 디지털데일리: 에이전트AI, 생성AI보다 전력 소모 ‘136.5배’↑…”에너지 인프라 중요성 부각” (2026.7.5)
- 한국일보: AI 에이전트는 전기 먹는 하마? “생성형 AI보다 136배 더 써” (2026.7.5)
- 한겨레: ‘답변’ 아닌 ‘과제’ 수행하는 AI, 에너지 소비 136.6배 많아 (2026.7.5)
