A single complex request to an advanced AI agent consumes an average of 348.41 watt-hours of electricity, up to 136.5 times more than a simple question answered by a conventional chatbot, according to a new study from the Korea Advanced Institute of Science and Technology (KAIST).
Published at the 2026 IEEE International Symposium on High Performance Computer Architecture (HPCA), the paper provides a system-level measurement of the hidden energy costs of AI agents.
AI agents differ from standard chatbots because they independently plan, call external tools such as search engines and calculators, and loop through their reasoning repeatedly to complete a task, says the paper. The KAIST team found that this process can make response times up to 153.7 times longer. Crucially, expensive graphics processing units (GPUs) sit idle for as much as 54.5% of total execution time, burning electricity while waiting for external websites and applications to respond.
The researchers, led by Professor Rhu Min-soo, estimated that if AI agents become mainstream and handle 13.7 billion requests globally per day, total data centre power demand would reach approximately 198.9 gigawatts, roughly half of the average total electricity consumption of the entire United States. The current power grid, they warned, cannot sustain such a load.
“This is the first case to quantitatively present how much power and cost are required not just to make AI smarter, but to implement and sustain intelligence,” Professor Rhu said. He called for a “co-design” approach that jointly optimises AI models, semiconductors, data centres, and power infrastructure to avoid a looming sustainability crisis. The research team has released the agent implementation techniques and benchmarking environment as open source for further study.