KAIST study warns AI agents may demand half of U.S. electricity supply

Researchers at the Korea Advanced Institute of Science and Technology (KAIST) have published the first comprehensive analysis of the energy cost of AI agents—systems that autonomously plan, reason and execute multi‑step tasks. Measuring real‑world workloads, they found that an AI agent using a 70‑billion‑parameter language model consumes about 348.41 watt‑hours per query, roughly 136.5 times more energy than a conventional generative‑AI chat response. Response latency can be up to 153.7 times longer, and GPUs remain idle for about 54.5 percent of the execution time while waiting for external tools.

When the team projected a future in which agents handle 13.7 billion requests per day—comparable to today’s Google search traffic—the required power would be about 198.9 GW, roughly half of the United States’ average electricity consumption. The authors argue that further AI progress will depend not only on larger models but also on more efficient semiconductors, better GPU utilisation, data‑center design and power‑grid capacity, urging a co‑design approach for AI hardware and infrastructure.

The study also notes the rapid proliferation of AI agents, citing hundreds of thousands of verified agents on platforms such as Molbook and stable‑coin‑enabled services, underscoring the urgency of addressing the hidden electricity costs as agentic AI becomes mainstream.