AI Agents' Soaring Energy Consumption Raises Market Concerns: KAIST Warns

A groundbreaking study published by researchers at the Korea Advanced Institute of Science and Technology (KAIST) has highlighted the substantially higher energy consumption of autonomous artificial intelligence (AI) agents compared to simpler chatbots. The team, led by Professor Minsoo Rhu, warned that AI agents could consume up to 136.5 times more electricity per query, posing significant operational and environmental costs for data centers globally.

According to the research, AI agents, when executing complex tasks, repeatedly invoke large language models, conduct web searches, perform calculations, and coordinate multiple steps. This iterative process can extend response times by up to 153.7 times and lead to expensive Graphics Processing Units (GPUs) sitting idle for as much as 54.5% of their total execution time while awaiting external tools. This significant idle time still contributes to energy consumption, raising serious questions about the overall efficiency of AI agents.

The KAIST team found that a single query processed by an AI agent using a 70-billion-parameter large language model consumed an average of 348.41 watt-hours (Wh) of electricity. This is vastly more than the energy expended by conventional generative AI chatbots for a simple question-and-answer interaction. In a simulated scenario of widespread AI agent adoption, assuming 13.7 billion daily requests, the projected data center power demand could reach approximately 198.9 gigawatts (GW), roughly equivalent to half of the total electricity consumption of the United States.

These findings suggest that competitiveness in the AI era will extend beyond mere model performance to encompass the efficiency of data centers and power infrastructure. The escalating demand for AI is already driving unprecedented investment in data centers, surpassing other commercial building construction. The International Energy Agency (IEA) and other bodies forecast that global electricity consumption by data centers could double or triple by 2028-2030. This places additional strain on energy grids, heightening concerns about infrastructure capacity and sustainability.

Analysts and market experts emphasize that a new approach is essential for the sustainable development of AI. The center of gravity in AI competitiveness is expected to shift from developing smarter software to an integrated co-design approach, optimizing AI models, chips, and data center infrastructure with a strong focus on energy efficiency. This marks a critical area where technology companies and energy providers will need to re-evaluate and reshape their strategies in the coming period. It is noted that existing energy forecasts may not adequately account for the 'autonomy multiplier' of AI agents, potentially severely underestimating future power demands.