AI agents consume up to 136 times more energy than typical generated AI.
A research team at
AI Agents Consume 136 Times More Power Than Standard Generative AI - Seoul Economic Daily
https://en.sedaily.com/technology/2026/07/06/ai-agents-consume-136-times-more-power-than-standard
Large-scale language models (LLMs) like ChatGPT primarily serve to answer user questions. On the other hand, AI agents can autonomously execute complex tasks, planning and utilizing external tools, enabling them to solve more intricate problems in various fields such as software development, research support, and task automation.
A research team led by Professor Yoo Min-soo of the Department of Electrical and Electronic Engineering at KAIST analyzed the computational load and energy consumption generated during the execution of these AI agents. According to the research team, while the use of AI agents is expanding, the processing power and costs required to operate AI agents as actual services are still not fully understood, and this is the first time that the computational costs and power consumption of AI agents have been quantitatively analyzed.
The analysis revealed that the AI agent invoked LLM an average of 9.2 times more often than a generative AI performing normal stepwise inference. Response times increased by up to 153.7 times, and the time the GPU spent idle while using external tools accounted for 54.5% of the total execution time. This indicates that the AI agent was not fully utilizing its hardware, as the GPU spent more time idle than performing calculations.
The increase in power consumption was even more pronounced: an AI agent using a 70 billion-parameter LLM, a scale also used in commercial AI services, consumed an average of 348.41 Wh of power to process a single query. According to the research team, this represents up to 136.5 times the power consumption of a typical generative AI.
The research team also estimated the power demand if the entire data center were operational. Assuming that AI agent requests reach 13.7 billion per day in the future, the data center would require approximately 198.9 GW of power. This is estimated to be about half of the average power consumption in the entire United States.
Professor Yu stated, 'This research goes beyond simply making AI smarter; it's the first to quantitatively demonstrate how much power and cost are required to realize and maintain that intelligence. In an era where AI agents are widespread, an approach that integrates and optimizes not only AI data centers but also AI agent models and power infrastructure will be crucial.'
It should be noted that the estimate of 'consuming up to 136.5 times more energy' is based on a comparison between 'simple question-and-answer sessions in generative AI' and 'complex task processing by AI agents,' and the figure of '136.5 times' is open to interpretation. Choi Ki-young, former professor at Seoul National University and former Minister of Science and ICT, pointed out, 'While it is difficult to directly compare generative AI and AI agents, it is clear that solving complex problems requires far more energy than simple chat. The KAIST paper is important because it highlights that simply increasing the size and cost of AI data centers may not be sustainable in order to meet the current demand for AI agents.'
Related Posts:
in AI, Posted by log1e_dh