KAIST announced on July 5, 2026 that a team led by chair professor Yoo Min-soo of its School of Electrical Engineering has, for the first time in the world, quantified how an AI agent consumes an average of 348.41 watt-hours to handle a single query, which is 136.5 times more power than a conventional generative-AI question-and-answer exchange. The team measured an agent built on a 70-billion-parameter large language model in a real service setting, and presented the work at the HPCA conference in February 2026. This article organizes the primary figures from KAIST's announcement and Korean reporting, and examines what those numbers actually mean.

Where the 348.41-Watt-Hour Figure Comes From

The object of measurement was not a simple chatbot but an AI agent that repeatedly calls tools to carry a task forward. The team measured that an agent running a 70-billion-parameter large language model consumes an average of 348.41 watt-hours per query, which is 136.5 times higher than using the same model for plain question-and-answer. The point is usage pattern, not model size. When a single query stacks up searches, calculations, and external tool calls, the structure that runs the same network over and over drives power up.

Power Leaks From Waiting, Not Computing

Agents burn power mainly through idle waiting for tools to finish, not through heavy computation itself. The team found that an agent's response time stretched up to 153.7 times longer than a Chain-of-Thought approach, and that for much of that time the GPU sat idle up to 54.5 percent while it waited for external tools to finish. In other words, an expensive accelerator stays powered on while doing nothing for long stretches. This diagnosis, that the main culprit is idle waiting rather than heavy computation itself, points to the structure of execution rather than model slimming as the place to improve efficiency.

How to Read the 198.9-Gigawatt Projection

The most striking number is a future scenario. Simulating a world with 13.7 billion agent requests per day, the team estimated data-center power demand at about 198.9 gigawatts, which Korean reporting framed as roughly half of the average total power consumption of the entire United States. What is impressive is less the sheer magnitude than the fact that the study pins down, in numbers, how power infrastructure rather than performance could become the ceiling on agent adoption. That said, this figure is an estimate premised on a specific model and request volume, so it should be read as heavily dependent on the model size and tool configuration used in a real service.

What Teams Actually Running Agents Should Take Away

The message for practitioners is to look at the flip side of the performance race. A team putting an agent into a product should factor per-query power and latency into design as operating costs, not just answer quality. If idle GPUs are a large share of the cost, parallelizing tool calls or scheduling waiting accelerators onto other requests can translate directly into savings. With data-center power already a bottleneck for siting and expansion in Korea too, the direction worth noting is that efficiency, running more agents on the same power rather than only scaling models up, becomes a real competitive edge.

Why the Numbers Deserve Careful Reading

The study's values should be read as measurements and estimates from a single configuration. The 348.41 watt-hours and 136.5x figures come from running a 70-billion-parameter model in one specific agent mode, and the gap could narrow with smaller models or optimized execution. The 198.9-gigawatt scenario likewise rests on the assumption of 13.7 billion daily requests and rises or falls with actual adoption. Even so, the contribution lies less in the precision of any one number than in being the first to quantify, in a real environment, the agent power cost that performance metrics had obscured, and to release the measurement tools as well.

Source: Herald Corp · Digital Today