A study found that if "AI agents," which plan tasks on their own and even use tools, become widespread, the power burden on data centers could surge sharply. Unlike Generative AI that simply answers questions, AI agents repeat multiple rounds of decision-making, significantly increasing response times and power consumption.
Korea Advanced Institute of Science and Technology (KAIST) said on the 5th that a research team led by endowed professor Yu Min-su of the School of Electrical Engineering analyzed AI agents' use of computing resources and energy consumption under assumed real service conditions. The team viewed AI agents not as a single program but as a new type of workload that data center servers and GPUs must process continuously, and examined performance degradation factors alongside power draw.
The analysis found that AI agents did not stop at generating a single response like existing Generative AI. To solve problems, they created plans, repeatedly requested needed information, and incorporated results from external tools. As a result, calls to the large language model increased, and total response time was up to 153.7 times longer than the conventional approach.
GPU utilization efficiency also fell. While external tools such as search or code execution carried out tasks, the GPU could not compute for a substantial period. The team said the share of total run time during which the GPU remained idle reached as high as 54.5%. This means that even if costly AI Semiconductor chips are deployed, there is considerable downtime in actual processing.
The gap in power use was even more pronounced. Using a 70-billion-parameter large language model—on par with current commercial AI services—as the baseline, an AI agent consumed an average of 348.41 watt-hours (Wh) per request. That is up to 136.5 times more than typical Generative AI question-and-answer tasks.
The team also estimated a scenario in which AI agent requests grow to 13.7 billion per day. In that case, the required data center power was calculated at about 198.9 gigawatts (GW). This far exceeds the scale of multi-gigawatt AI data centers that countries have recently been building.
The researchers viewed the findings as indicating that the AI industry's focus could shift from model performance competition to energy efficiency competition. They said that going forward, there are limits to optimizing AI models, semiconductors, servers, cooling systems, and power grids separately, and that a strategy to design the entire infrastructure together is needed.