AI Agents: Energy Consumption Up to 136.5 Times Higher According to New Study

The Korean University KAIST has measured for the first time the hidden energy costs of AI agents. In particular, it highlights how these systems can consume up to 136.5 times more energy per request compared to traditional generative AI. The study, authored by Professor Minsoo Rhu's group at the School of Electrical Engineering, shifts the focus from model accuracy to the efficiency of data centers and electrical infrastructures.

Key Data

In the paper presented in February at HPCA 2026, KAIST analyzed the behavior of AI agents in a realistic context, viewing them not just as software but as a workload impacting servers and GPUs. The most evident result is an average consumption of 348.41 Wh per query on a 70 billion parameter LLM, compared to much lower values from classic generative systems.

The point is not just the final wattage, but how these systems operate: they call the language model multiple times, use external tools, and multiply the steps needed to arrive at an answer. KAIST points out that this architecture leads to latencies up to 153.7 times higher and leaves GPUs idle for up to 54.5% of the time, while external tools complete their tasks.

Why It Matters So Much

AI agents are designed to plan, search, use calculators, execute code, and compose multiple autonomous actions. This chain of operations makes it harder to maintain high GPU usage levels, which remain idle waiting for responses from other tools, ultimately wasting a significant portion of execution time. For these reasons, when AI agents come into play, the role of CPUs as orchestrators of GPU work becomes crucial.

KAIST views the phenomenon as a co-design problem: it’s not just about improving the model; it requires more efficient chips, data centers, and electrical networks to work together. In other words, the competition in AI no longer revolves solely around who produces the most capable system but also around who can make it sustainable in terms of energy and operations.

Future Scenario

The study also attempts to push the reasoning on an industrial scale. If AI agents were to reach 13.7 billion requests a day, a volume comparable to Google searches, the demand from data centers would touch approximately 198.9 GW, a threshold well beyond the scale of the AI hubs currently under construction.

For this reason, KAIST emphasizes an integrated approach, where model architecture, computational infrastructures, and electrical power are designed together. This is the step that, according to researchers, can keep operational costs and energy impact under control as AI agents become more widespread.

As for KAIST, short for Korea Advanced Institute of Science and Technology, it is one of the leading research universities in South Korea. Founded in 1971 in Daejeon, it focuses its activities on science, engineering, and technology, playing a central role in advanced research development and collaborations with industry and international institutions.