The next generation of artificial intelligence could trigger a massive global energy crisis.
A groundbreaking new study by the Korea Advanced Institute of Science and Technology (KAIST) has quantified the hidden electricity costs of advanced AI agents for the first time. The findings reveal that these autonomous systems consume up to 136.5 times more energy per query than the conventional generative AI tools we use today.
While standard chatbots simply answer a question and stop, AI agents act like digital assistants. Given a goal like planning a vacation or managing a budget, they independently figure out how to do it. To finish the job, the agent will search the internet, make calculations and execute commands entirely on its own.
However, this independence comes at a massive environmental cost.
To complete a complex task, the agent has to talk to itself and rerun its core AI programming over and over again. This continuous looping means answers can take 153.7 times longer to generate. Even worse, expensive computer graphics chips sit completely idle more than half the time, burning electricity while simply waiting for external websites and apps to respond.
The strain on global data centers is immense.
The KAIST research team, led by Distinguished Professor Minsoo Yoo of the School of Electrical Engineering, found that a single complex request to an AI agent burns through an average of 348.41 watt-hours of electricity. If these agents become mainstream and handle an estimated 13.7 billion requests globally per day, the power grid will not be able to cope. Total data center electricity demand would skyrocket to approximately 198.9 gigawatts — roughly half of the average energy consumption of the entire United States.
KAIST announced on July 5 that the team has, for the first time in the world, quantitatively analyzed how much computational resources and power AI agents use under real service conditions. The research results were presented in February at the 32nd IEEE International Symposium on High-Performance Computer Architecture (HPCA).
The analysis showed that AI agents compute differently from conventional generative AI. They repeatedly invoke large language models while using multiple external tools such as internet search, calculators, and code execution. Response times increased up to 153.7 times, and GPUs waited idle without computation for up to 54.5% of total execution time.
An AI agent based on a 70-billion-parameter LLM consumed an average of 348.41 watt-hours per question — up to 136.5 times higher than simple question-and-answer generative AI.
Professor Yoo said competitiveness in the AI era is expanding from “smarter AI” to “more efficient AI,” and that co-design is needed to jointly optimize AI models, semiconductors, data centers, and power infrastructure.
The research team released its AI agent implementations and benchmarking environment as open source at https://github.com/VIA-Research/AgentBench.
Paper title: “The Cost of Dynamic Reasoning: Demystifying AI Agents and Test-Time Scaling from an AI Infrastructure Perspective” (doi.org/10.1109/HPCA68181.2026.11408569).