Agentic AI could consume 136× more energy than chatbots, study warns
Agentic AI could consume up to 136.5× more energy per query than standard generative AI, according to new research from KAIST. Unlike a simple call‑and‑response chatbot, an AI agent repeatedly pings its model as it reasons through multi‑step tasks, creating a massive multiplier effect in energy use and latency. The study found an agent can require 348.41 Wh per query—roughly the energy needed to keep an LED bulb lit for 24 hours—and may take 153.7× longer to respond. GPUs sit idle for over 54% of the time during agentic tasks, creating huge inefficiencies.
This matters because agentic AI is already spreading fast: over 200,000 verified agents exist on Moltbook, and 400,000 have been approved to use USDC. Google and others are integrating agents directly into consumer products. If agents ever reach Google‑scale query volumes (13.7 billion/day), global demand could hit 198.9 GW—about half of the United States’ total electricity consumption. Without major efficiency breakthroughs, agentic AI could become one of the most energy‑intensive technologies ever deployed.