LLevelNews
environmentAdvanced2 min read7/8/2026

KAIST study finds AI agents are far more energy-intensive than chatbots

Researchers at KAIST measured the power use of AI agents and found large gaps when compared with chatbots. They warn that wider use could strain data centres, chips and electricity networks.

As technology companies shift attention from chatbots such as ChatGPT, Gemini and Grok towards AI agents, systems that can carry out specific tasks on their own, a new KAIST study points to a major cost: electricity. The research is among the first to measure AI agent power consumption with concrete figures, and it suggests that the move from one-shot chat responses to multi-step autonomous work can be extremely energy-intensive. In tests on a 70-billion-parameter large language model close to commercial scale, a single complex agent query used an average of 348.41 watt-hours. The Reflexion framework, running on Meta’s Llama-3.1-Instruct 70B model, consumed 136.5 times more GPU energy per query than a standard chatbot, while LATS reached 62.1 times that level. The study also found that response times could be up to 153.7 times longer, largely because GPUs spent much of their time waiting for external websites or software tools rather than computing. The researchers project that if daily global use of AI agents reaches about 13.7 billion requests, total data-centre power demand could rise to around 199 gigawatts, roughly half of average electricity use in the United States. Professor Rhu Min-soo said software improvements alone are unlikely to solve the problem, and the team argues that AI models, microchips and data-centre electricity systems may all need redesigning.

More stories to read

We use Google Analytics cookies to understand how the site is used. They stay off unless you accept. Privacy Policy