When It Comes to Energy Use, AI Agents Could Make Chatbots Look Like Pocket Calculators - Gizmodo

2026-07-06 · Source: artifical intelligence via Google News · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Novice, quick

Summary

A new research paper from the Korea Advanced Institute of Science and Technology quantifies the "hidden costs" of AI agents, revealing they consume up to 136.5 times more energy per query than standard generative AI models. This increased demand, averaging 348.41 watt-hours per query (equivalent to an LED light bulb on for a full day), stems from agents' multi-step reasoning, which requires continuous model interaction. Beyond energy, agents exhibit 153.7 times longer response latency and cause GPUs to sit idle for up to 54.5% of the time. With 200,000 agents on Moltbook and 400,000 approved for UDSC, and companies like Google integrating them, their prevalence is growing. Projections suggest 13.7 billion daily agent requests could demand 198.9 gigawatts, nearly half of the United States' current electricity consumption.

Key takeaway

For AI architects and infrastructure planners evaluating agentic AI deployments, it is critical to account for their substantially higher energy and resource demands. Your current scaling models for generative AI will likely underestimate the computational load, latency, and GPU idle times associated with agents. Prioritize efficiency optimizations and re-evaluate infrastructure strategies to mitigate the significant environmental and operational costs before widespread adoption.

Key insights

AI agents significantly increase energy and resource consumption due to their iterative, multi-step reasoning processes.

Principles

• Agentic AI's multi-step reasoning drives higher energy demand.
• Continuous model interaction increases latency and GPU idle time.

In practice

• Quantify agent energy costs before deployment.
• Optimize agent workflows to reduce model pings.
• Consider latency impact on resource utilization.

Topics

• AI Agents
• Energy Consumption
• Large Language Models
• Resource Efficiency
• Computational Latency
• AI Infrastructure

Best for: MLOps Engineer, AI Product Manager, Product Manager, Tech Journalist, General Interest, AI Architect

Related on AIssential

• Towards Understanding, Analyzing, and Optimizing Agentic AI Execution: A CPU-Centric Perspective — CPU-centric bottlenecks in agentic AI workloads significantly impact latency, throughput, and energy consumption, often more than GPU inference.
• Being friendly to your AI might be the least eco-friendly thing you can do — Inefficient AI interaction patterns, like verbose prompts, significantly increase energy consumption at scale.
• TAI #197: Anthropic Turned the OpenClaw Demand Signal Into a Product — Agentic AI advancements are rapidly increasing compute demand, straining a constrained global AI supply chain.
• not much happened today — AI agents are expanding beyond APIs to direct computer control, while model training and evaluation become more efficient and industrialized.
• AI Wearable Watches You Back — AI's future hinges on co-designed, continuously running agentic systems requiring vastly increased, energy-efficient compute.
• The Real Costs of AI Agents Doing Human Jobs — The "Agents Last Exam" reveals current AI agents struggle with industrial tasks, highlighting performance and cost disparities.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by artifical intelligence via Google News.