AI agents consume 136.5 times more power than ChatGPT, KAIST study finds

An AI-generated image illustrating research into AI agent power consumption. [Provided by KAIST]

South Korean researchers have found that AI agents consume up to 136.5 times more energy per query than conventional generative AI systems.

A research team led by distinguished professor Yoo Min-soo of KAIST's School of Electrical Engineering announced Saturday that it had conducted what it described as the world's first systematic analysis of the computational resources and power consumed by AI agents operating in real service environments.

Large language model-based applications such as ChatGPT have rapidly evolved beyond simply answering questions. They now function as AI agents capable of independently forming plans and solving complex problems by drawing on external tools such as internet search, calculators and code execution.

AI agents are seeing expanding use across software development, research and workplace automation, yet the power and cost required to run them in production have remained largely unknown.

The research team defined AI agents not as simple programs but as a new class of workload — the total set of computational tasks a computer must carry out — that requires servers and GPUs (graphics processing units, the high-performance semiconductors that handle large-scale AI computation) in data centers to process continuously. The team then analyzed the computational load and energy consumption generated during actual execution.

The analysis found that, unlike conventional step-by-step inference, AI agents repeatedly call large language models multiple times throughout a single task.

Those repeated model calls also dramatically increased response times — by as much as 153.7 times. Meanwhile, GPUs sat idle for up to 54.5 percent of total execution time while external tools carried out their work. In other words, the more complex the task an AI agent performs, the less efficiently it uses expensive GPU hardware, creating a new form of inefficiency.

The KAIST research team behind the study (from left): Shin Byeong-jun, master's student; Jeong Jin-ha, integrated master's-doctoral student; Kim Ji-in, doctoral student; and Professor Yoo Min-soo. [Provided by KAIST]

The research team also analyzed AI agent power consumption at the data center scale. An AI agent running a large language model with 70 billion parameters — the values that store an AI's learned knowledge and capabilities, representative of current commercial AI services — consumed an average of 348.41 watt-hours of electricity per query. That is 136.5 times higher than the energy used by conventional generative AI in simple question-and-answer interactions.

Projecting a future scenario in which 13.7 billion AI agent requests are processed daily, the team estimated that data center power demand would reach approximately 198.9 gigawatts. That figure far exceeds the multi-gigawatt AI data centers currently being built around the world and is equivalent to roughly half of the United States' total average power consumption.

The findings are expected to become foundational to efforts to reduce the operating costs of AI services and build sustainable AI infrastructure.

"This research demonstrates that improvements in AI agent performance go beyond simply demanding more computation — they impose an entirely new burden on data center infrastructure as a whole," Yoo said. "Research and investment aimed at dramatically reducing the cost borne by end users of AI services, while at the same time building sustainable AI infrastructure, are essential."

nbgkoo@heraldcorp.com