![\"Power](\"https:\/\/www.europesays.com\/ai\/wp-content\/uploads\/2026\/07\/32e8d7b395f9d899845949febc8abe67.jpg\"\/)
<\/p>\n<\/p>\n
A Korean research team has quantitatively analyzed, for the first time in the world, the power consumption and efficiency of AI agents. Provided by Getty Images Bank.<\/p>\n
Analysis has found that an \u201cAI agent,\u201d an artificial intelligence (AI) system that independently plans and uses external tools such as internet search and code execution to solve problems, consumes up to 136 times more power per question than a simple question-and-answer generative AI. The study argues that beyond competition over AI model performance, the efficiency of data centers and power infrastructure must be considered together.<\/p>\n
KAIST announced on the 5th that the team led by Distinguished Professor Minsoo Yoo of the School of Electrical Engineering has, for the first time in the world, quantitatively analyzed how much computational resources and power AI agents use under real service conditions. The research results were presented last February at the 32nd IEEE HPCA, an international conference in computer system architecture.<\/p>\n
Recently, AI has evolved beyond simply answering questions into AI agents that independently plan and use various external tools such as internet search, calculators, and code execution to solve complex problems. While their applications are expanding in areas such as software development, research, and task automation, the power and computational costs required for actual service operations have not been known.<\/p>\n
The research team defined AI agents as a new type of workload that data center servers and graphics processing units (GPUs) must handle, and analyzed the computational load and energy consumption that occur during actual execution.<\/p>\n
The analysis showed that AI agents compute in a different way from existing generative AI, which derives answers by unfolding a chain of thought step by step like a human. They repeatedly invoke large language models (LLMs) while using multiple external tools.<\/p>\n
In this process, response time increased significantly. The response time of AI agents increased by up to 153.7 times, and it was found that while external tools were performing tasks, GPUs waited idle without computation for up to 54.5% of the total execution time. This means that the more complex tasks AI performs, the less efficiently it utilizes expensive GPUs.<\/p>\n
Power consumption also increased sharply. According to the research team, an AI agent based on a large language model with 70 billion parameters used an average of 348.41 watt-hours (Wh) to process a single question. A watt-hour is a unit of energy that indicates how much electricity is used, representing the amount of electrical energy consumed when using 1 watt (W) of power for 1 hour. This is up to 136.5 times higher than the simple question-and-answer mode of conventional generative AI.<\/p>\n
Assuming a future environment in which 13.7 billion AI agent requests occur per day, the researchers also estimated data center power demand. They found that approximately 198.9 gigawatts (GW\u2014a large-scale power capacity used in national power grids) of power would be needed. This level far exceeds the scale of AI data centers currently being pursued by countries and corresponds to about half of the United States\u2019 average total power consumption.<\/p>\n
The research team predicts that future AI competitiveness will extend beyond model performance to include power efficiency. They explained that \u201cco-design,\u201d which jointly optimizes AI models, semiconductors, data centers, and power infrastructure, will be necessary.<\/p>\n
Distinguished Professor Minsoo Yoo said, \u201cThis is the first case to quantitatively present how much power and cost are required not just to make AI smarter, but to implement and sustain intelligence,\u201d adding, \u201cIn an era where AI agents become ubiquitous, research and investment in co-design will be essential to drastically reduce the cost for end users to access AI services and to build sustainable AI infrastructure.\u201d<\/p>\n
The research team has released as open source the AI agent implementation techniques and the benchmarking environment for AI performance evaluation used in the paper so that researchers around the world can use them for follow-up studies. (https:\/\/github.com\/VIA-Research\/AgentBench)<\/p>\n
doi.org\/10.1109\/HPCA68181.2026.11408569<\/p>\n
\u00a0<\/p>\n
![\"From](\"https:\/\/www.europesays.com\/ai\/wp-content\/uploads\/2026\/07\/6ebcbe7dbf7cd229b42ad845455d32b9.jpg\"\/)
<\/p>\n
From left: Master\u2019s student Byungjun Shin, integrated Master\u2019s-PhD student Jinha Jung, PhD student Jiin Kim. At the top, Professor Minsoo Yoo. Provided by KAIST.<\/p>\n
Copyright \u24d2 DongA Science. All rights reserved.<\/p>\n","protected":false},"excerpt":{"rendered":"A Korean research team has quantitatively analyzed, for the first time in the world, the power consumption and…\n","protected":false},"author":2,"featured_media":95429,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[6],"tags":[2069,24,405,25,7537,42881,85,7975,223,2010,1371,205,22771,1807,4835,4754,49874,12760,49875,49876,49877,49878,49879,12764,49880,49881,49882],"class_list":["post-95428","post","type-post","status-publish","format-standard","has-post-thumbnail","category-agentic-ai","tag-agent","tag-ai","tag-ai-agents","tag-artificial-intelligence","tag-artificial-intelligence-agents","tag-benchmark","tag-data-center","tag-efficiency","tag-generative-ai","tag-gpu","tag-graphics-processing-unit","tag-infrastructure","tag-kaist","tag-large-language-model","tag-power","tag-server","tag-49874","tag-12760","tag-49875","tag-49876","tag-49878","tag-49879","tag-12764","tag-49880","tag-49881","tag-49882"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/95428","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=95428"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/95428\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/95429"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=95428"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=95428"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=95428"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}