AI agents, the next generation of autonomous AI systems that can reason and act independently, consume dramatically more power than traditional AI applications. A research team at KAIST (Korea Advanced Institute of Science and Technology) has published the first comprehensive analysis of how much electricity these systems actually require, finding that a single AI agent query consumes 348.41 watt-hours of energy. This is 136.5 times higher than the energy consumed by conventional generative AI systems performing simple question-answering tasks.

What Exactly Are AI Agents and Why Do They Use So Much Power?

AI agents represent a significant leap beyond chatbots like ChatGPT. Rather than simply answering questions, these systems can plan multi-step tasks, use external tools like web search and code execution environments, and solve complex problems autonomously. Think of them as AI systems that don't just respond to a single prompt, but instead break down a goal into multiple steps and execute them independently.

The power consumption spike comes from how these systems operate. When an AI agent tackles a complex task, it repeatedly calls large language models (LLMs), which are the neural networks that power modern AI systems. A conventional AI system might process a single request and return an answer. An AI agent, by contrast, might call the language model dozens of times as it reasons through a problem, checks its work, and coordinates different tools. This constant back-and-forth with the GPU (graphics processing unit), the specialized chip that runs these calculations, creates massive energy demands.

How Much Idle Time Are Data Centers Wasting on AI Agents?

The KAIST research team uncovered another efficiency problem: GPUs sit idle for extended periods while AI agents wait for external tools to finish their work. The study found that GPUs remain unused for as much as 54.5 percent of the total execution time while the agent waits for web searches, calculations, or code to run. This means expensive hardware is essentially doing nothing, yet still consuming power and occupying valuable data center space.

Response times also suffer dramatically. While a conventional AI system might respond in seconds, AI agents can take up to 153.7 times longer to complete their tasks because they must coordinate multiple steps and wait for external processes to finish. This latency problem compounds the energy issue, as GPUs must remain powered and ready even during these idle periods.

What Would Happen If AI Agents Reached Google-Scale Traffic?

The KAIST team projected a scenario that reveals the scale of the challenge ahead. If AI agents were to handle 13.7 billion requests per day, a volume equivalent to current Google search traffic, data centers would need approximately 198.9 gigawatts of power. To put this in perspective, that's roughly half of the average power consumption of the entire United States. Current AI data centers under development operate at only a few gigawatts, meaning we would need infrastructure roughly 50 to 100 times larger than what exists today.

This projection underscores why the KAIST team believes the competitive landscape of AI is shifting fundamentally. The race is no longer just about building smarter AI models, but about building efficient ones that don't require impossible amounts of electricity to operate at scale.

How to Improve AI Agent Energy Efficiency

• Co-Design Approach: Optimize AI semiconductors, data center infrastructure, and power systems together rather than treating them as separate problems. This integrated strategy can reduce operational costs and environmental impact simultaneously.
• GPU Utilization Improvements: Develop scheduling and batching techniques that keep GPUs active during AI agent execution, reducing the 54.5 percent idle time currently observed in these workloads.
• Model and Infrastructure Joint Optimization: Redesign AI agent models themselves to require fewer language model invocations, while simultaneously upgrading data center power delivery and cooling systems to handle peak loads more efficiently.

"This study is the first to quantitatively show not only how AI is becoming more intelligent, but also how much electricity and cost are required to implement and sustain that intelligence. As AI agents become widespread, it will become increasingly important to take an integrated co-design approach that optimizes not only AI data-center infrastructure, but also AI agent models and power infrastructure," said Professor Minsoo Rhu.

Professor Minsoo Rhu, School of Electrical Engineering at KAIST

Why Should Tech Companies Care About These Findings?

The implications extend far beyond academic interest. As AI agents are increasingly adopted for software development, research, and workplace automation, the cost of running these systems will directly impact their commercial viability. A company deploying AI agents at scale could face electricity bills that dwarf the cost of the AI models themselves. This creates a powerful incentive for innovation in both hardware and software efficiency.

The research team, led by Professor Minsoo Rhu, presented their findings at the IEEE International Symposium on High-Performance Computer Architecture, one of the most prestigious conferences in computer system design. They have also released their AI agent implementations and benchmarks as open source, enabling researchers worldwide to build on their work and develop more efficient solutions.

The bottom line is clear: the next frontier of AI competition isn't just about model intelligence, but about sustainable, cost-effective infrastructure. Without significant advances in efficiency, AI agents risk becoming too expensive to operate at the scale that would make them truly transformative for businesses and society.