Language
Change font
A
A
A
A
Share
TOP
Most read
Language
Change font
A
A
A
A
Share
Artificial intelligence (AI) agents have been shown to be significantly less energy efficient than conventional chatbot-type AI. The AI agent needed up to 136.5 times more power to handle one question. This is why it is necessary to consider expanding power infrastructure together when planning AI industries or policies in the future.
Yoo Min-soo, chair professor of electrical and electronic engineering at KAIST, said on the 5th that he analyzed how much computational resources and power AI agents use. This is the first quantitative analysis of energy use by AI agents in the world.
Recently, AI is gradually increasing its accuracy, but its efficiency is rapidly deteriorating. As the amount of calculation increases, the accuracy increases, but the increase gradually decreases. The efficiency of AI agents in which several AI models cooperate is also inevitably reduced. The researchers evaluated the overall consumption of resources, including power consumption and graphics processing unit (GPU) usage of five previously released AI agents.
Studies have shown that AI agents have longer response times than conventional AI models and lower GPU efficiency. Answer times are increased up to 153.7x, and GPUs are shown to wait up to 54.5% of the total run time without calculating anything. In other words, the more complicated AI does, the less fully it utilizes GPUs.
The overall inefficiency is increasing because AI agents call LLM several times to utilize it. While Generative AI used LLM once for one question, one of the AI agents analyzed this time used an average of 71 times. In this case, in about 70 gaps without using LLM, the GPU waits without any separate work. The quality of the answer is improved, but the efficiency is reduced.
AI agents spent an average of 348.41 watt-hours (Wh) processing one question. It is up to 136.5 times higher than the existing Generative AI. Currently, there are about 13.7 billion Google searches a day, and if all of the same searches are made with AI agents, about 198.9 gigawatts (GW) are needed. It is equivalent to about half of the total electricity consumption in the United States.
The researchers suggest that the front lines of the AI industry should change from accuracy to efficiency in the future. Instead of recklessly increasing the accuracy, it is necessary to think about how to achieve the same accuracy with fewer calculations. As resources cannot be invested endlessly in AI, the challenge is to find an appropriate relationship between accuracy and efficiency.
Specifically, it is said that ‘calculated cognitive reasoning’ is needed to mix AI models of various sizes within AI agents and weigh how much calculation is needed before answering questions. Don’t use a large AI model for all questions, but simply answer with a small AI model if it’s a problem that can be calculated a little.
Professor Yoo said, “In the future, when AI agents become common, an optimization approach that co-designs not only AI data center infrastructure but also power infrastructure will become more important.”
2026-07-04 21:40:54
2026-07-05 15:30:35
2026-07-04 06:03:14
2026-07-05 16:23:27
2026-07-04 13:07:02
2026-07-05 13:08:28
2026-07-05 06:01:10
2026-07-05 11:33:00
2026-07-05 13:29:44
2026-07-05 15:44:23
※ This article was translated using AI technology for reader convenience.
Maeil Business(MK) provides these translations “as they are” and makes no warranties of any kind, either explicitly or implicitly, regarding accuracy, reliability and marketability, suitability for a particular purpose, etc. of translation. Please be informed that the content provided may not be translated accurately due to limitations in machine translation before using this service.
Copyright (c) 매경AX. Maeil Business News Korea & mk.co.kr, All rights reserved.
Prohibition of unauthorized reproduction, redistribution, and use of AI learning