Much has been made about the emergence of generative AI and its strain on the power grid due to energy demand. So wait until you see how much energy the agent AI consumes. A new research paper from the Korea Institute of Science and Technology aims to quantify the “hidden costs” of AI agents, finding that they can consume up to 136.5 times more energy per query than generative AI models.
There is logic in the fact that AI agents require more processing power and energy than standard generative AI queries. LLM requests are typically call-and-response. A person enters a query and the model responds. However, agent AI typically requires multiple steps to execute a command. According to the researchers, this requires the agent to continually ping the model and generate new responses while reasoning through every step of a given task.
As a result, a synergistic effect occurs. According to the researchers, an AI agent running on a large language model, on the scale of most commercially available AI models, consumes an average of 348.41 Watt-hours per query. This is about the same as keeping an LED light bulb lit all day. They say this number is about 136.5 times the energy consumed by generative AI queries.
The impact of agent AI goes beyond energy consumption. The paper also investigated response latency and found that agent AI can take 153.7 times longer than standard queries. This is important because long response times strain computing resources. When an agent repeatedly pings a model to complete a task, the GPU can spend more time waiting than working. The researchers estimate that the GPU can be idle up to 54.5% of the time while the agent is performing a task, creating a level of inefficiency that doesn’t exist in simpler AI uses.
Now, all of this would be one thing if agent AI were just a concept being played around with in a lab somewhere. The reality is that we are already flooded with AI agents. At this point, we don’t really know how many agents are active. Moltbook, a social network for AI agents, reportedly has 200,000 verified agents registered and around 400,000 agents approved to use the stablecoin UDSC. Companies like Google are starting to incorporate agent AI into the web browsing experience. It is already quite popular.
The researchers also modeled a future in which AI agents generate 13.7 billion requests per day. This is about the same amount of queries that Google Search currently handles. They estimate that without significant improvements in energy efficiency, electricity demand will be around 198.9 gigawatts, roughly half of the current electricity consumption across the United States. I don’t know if this planet can handle half of America. But maybe we’ll find out.
