AI Agents Consume 136x More Energy Than Standard Chatbots

Energy demands multiply with multi-step AI reasoning

AI agents—systems that autonomously complete multi-step tasks rather than simply responding to prompts—consume dramatically more energy than conventional chatbots, according to new research from the Korea Advanced Institute of Science and Technology. The study quantifies what researchers call the "hidden costs" of agentic AI, finding these systems use 136.5 times more energy per query than standard generative AI models.

The research, first reported by Gizmodo, measured an AI agent running on commercial-scale language models consuming an average of 348.41 watt-hours per query—roughly equivalent to powering an LED bulb for 24 hours. By contrast, a typical chatbot query requires a fraction of that power for its single call-and-response interaction.

The energy gap stems from how agents operate. Unlike chatbots that generate one response per prompt, agents must repeatedly query their underlying models as they reason through each step of a complex task. This creates a multiplier effect: more model calls mean exponentially higher energy consumption.

Infrastructure inefficiency compounds the problem

Beyond raw energy use, the research identified severe inefficiency in how agents utilize computing resources. Response latency for agentic AI runs 153.7 times longer than standard queries, keeping GPUs occupied for extended periods. More troubling, the researchers estimate GPUs sit idle for 54.5% of the time while agents execute tasks—waiting between reasoning steps rather than actively processing.

This inefficiency matters because it ties up expensive hardware that could otherwise handle additional workloads. The idle time represents wasted capacity in data centers already straining to meet AI demand.

Why it matters

Agentic AI is already deployed at scale, not confined to research labs. Moltbook, a social network for AI agents, hosts 200,000 verified agents. Approximately 400,000 agents have been approved to use the USDC stablecoin. Google has begun integrating agentic capabilities into web browsing. Yet most organizations adopting these systems lack visibility into their true infrastructure costs.

The researchers modeled a scenario where AI agents handle 13.7 billion daily requests—matching current Google Search volume. Without major efficiency improvements, this would require 198.9 gigawatts of power, roughly half of total U.S. electricity consumption. As enterprises rush to deploy agents for email management, customer service, and workflow automation, the cumulative energy impact could strain electrical grids already challenged by AI data center growth.

The findings suggest organizations need to weigh the convenience of autonomous AI against substantial increases in operational costs and environmental impact. The gap between chatbot and agent energy consumption is not incremental—it represents a fundamental shift in infrastructure requirements.

These details were first reported by AJ Dellinger at Gizmodo.