Anime, manga, and games, with a take · A Yukimedia publication

← all stories other 1 sources · 1h ago ·

AI Agents Use Up to 136 Times More Energy Than Standard Generative AI, KAIST Study Finds

The study provides the first quantitative evidence that the computational and energy costs of running AI agents at scale may be unsustainably high, challenging the assumption that simply expanding data center capacity can meet future demand.

Reporting from 1 source: GIGAZINE.

AI Agents Use Up to 136 Times More Energy Than Standard Generative AI, KAIST Study Finds

A KAIST research team conducted the first systematic analysis of AI agent energy consumption in real service environments. They found that AI agents consume up to 136.5 times more power per query than standard generative AI, with GPUs idle 54.5% of execution time. The team estimated that widespread AI agent use could require 198.9 GW of data center power, roughly half of total US consumption.

Researchers at KAIST have published the first systematic analysis of how much energy AI agents consume in real service environments, and the numbers are stark. The team, led by Professor Yoo Min-soo, found that AI agents use up to 136.5 times more power per query than standard generative AI like ChatGPT. The agents call large language models an average of 9.2 times more often per task, and response times increase by up to 153.7 times. GPUs sit idle 54.5% of the time while agents use external tools, meaning the hardware spends more time waiting than computing.

For a single query using a 70-billion-parameter model, an AI agent consumed 348.41 Wh on average. If AI agent requests reach 13.7 billion per day, data centers would need about 198.9 GW of power, roughly half the average total electricity consumption of the United States. Former Minister of Science and ICT Choi Gi-young noted that while direct comparison between generative AI and AI agents is difficult, the paper clearly shows that solving complex problems requires far more energy than simple chat, and that simply scaling up data centers may not be sustainable.

Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.

Sources