South Korean Researchers: AI Agents Consume 136 Times More Power Than a Chatbot

Researchers at the Korea Advanced Institute of Science and Technology (KAIST) published, on July 5, 2026, the first detailed comparison of energy consumption between conversational AI models and autonomous agents carrying out complex tasks. The results show that agentic AI, increasingly promoted as the next stage of the industry's development, carries a hidden energy cost that has so far received little attention.

The difference stems from how the two types of systems operate. A standard chatbot receives a question, generates an answer, and stops. An AI agent works differently: it independently plans the steps needed to achieve a goal, such as booking a trip or preparing a budget, searches the web, performs calculations, and executes commands without human involvement.

Why agents draw so much power

To complete a complex task, an agent has to repeatedly "talk to itself," re-running its computational core at every stage of planning and verifying results. This continuous loop, described by the KAIST researchers in their paper as a new type of data center workload, is the main source of the huge gap in energy use compared with a simple question-and-answer model.

The KAIST team was the first to measure this effect under controlled laboratory conditions, using a model on a scale comparable to currently available commercial AI services, namely 70 billion parameters. Earlier analyses of AI energy consumption focused mainly on training large models or on simple chatbot queries, overlooking the growing category of agentic systems.

The scale of the problem at mass deployment

The most alarming part is the team's forward-looking scenario. If AI agents become widespread enough to handle around 13.7 billion queries a day, a volume close to today's global traffic on Google Search, total data center energy demand would jump to a level comparable to half of the entire United States' electricity consumption.

According to the study's authors, such a surge would mean that current energy infrastructure and transmission grids could not meet demand without a substantial expansion of generating capacity, including power plants dedicated solely to data centers.

Context for the industry

The publication lands at a moment when major cloud providers and model developers, from Microsoft to Google, are heavily promoting agentic features as the next step in AI product development, while disclosing little data on these systems' actual energy use. Some commentators note that the "136 times more" figure is sometimes taken out of context, since an agent performs work equivalent to many separate chatbot queries, making a strict one-to-one comparison not entirely fair.

Even so, the study's authors stress that the lack of transparency around the real energy cost of agentic AI makes it harder for energy planners and regulators to prepare for the coming rise in demand. In Poland, the issue has practical weight, since domestic data centers running cloud and AI services already compete with the power grid for connections, and plans to expand generating capacity must account not only for today's load but also for how fast AI agent adoption is growing among businesses.

The KAIST findings could also influence how companies design their agentic products, for instance by cutting down on redundant planning iterations or shifting part of the computation to smaller, specialized models instead of invoking a full-scale language model every time.

Sources: Korea Times (koreatimes.co.kr), EurekAlert (eurekalert.org), Forbes (forbes.com)