KAIST researchers measured roughly 348 watt hours per agent query, up to 136x a single chatbot call, and traced the gap to the multi step loops agents use to finish a task.
A single query to an AI agent burns roughly 348 watt-hours of electricity, about as much as keeping an LED bulb on for a full day. That figure comes from a KAIST research team that ran test-time scaling experiments across commercial-scale language models and now serves as one of the first anchored measurements of agentic AI's energy cost. Their recent arXiv preprint puts the energy of one agent interaction at up to 136.5 times the energy of a single generative AI call.
A chatbot answers in a single pass: one prompt in, one completion out. An AI agent carries a task forward in steps. It plans, calls a tool, evaluates the result, and re-enters the model to plan the next move. Each of those calls is a separate inference, billed separately against the GPU, and the KAIST team treats the resulting loop as a multiplier on top of inference cost. The multiplier scales with the number of reasoning turns the agent takes to finish a job.
The 136.5x ratio is what the team measured at the worst end, where heavily-scaled agents run many reasoning turns to complete a hard task. A simpler agent that fetches today's weather and stops burns roughly the energy of a few chatbot calls. An agent that browses a site, transcribes what it finds, drafts a reply, and waits for user feedback runs into the tens or hundreds of model calls before it is done. Energy cost tracks step count, and step count tracks how the agent is wired together.
GPUs sat idle about 54.5% of the time during agent runs in the team's instrumentation. The model is not the only expensive resource; the orchestration layer between model calls (file reads, browser actions, human-paced pauses, retries) leaves the accelerator waiting. That idle time is built into how agents work today, and it is invisible to users who only see the final answer. Closing the gap is a deployment question, not a model-size question.
That idle fraction matters now because agent work is no longer a research demo. Major platforms have shipped browser agents, productivity assistants, and developer tools that run multi-step tasks, which means an energy measurement at this scale translates to millions of real queries, not just one. The paper gives operators and grid planners an anchored number to plan around. KAIST's release frames the gap as a real-world grid-planning problem, and the same finding has since been reported in Gizmodo, Digital Trends, and The Engineer. Two questions follow the measurement. First, whether efficiency work on agent runtimes (shorter loops, smarter cache reuse, tighter guardrails on how many steps an agent is allowed to take) can pull the average down without breaking capability. Second, whether data center operators will price the difference. The KAIST team has given the industry its first figure. What gets built on top of it is the next story.