The energy tax associated with AI was already worrying. Studies indicate that AI agents are over a hundred times more detrimental.

      The increasing electricity demand from the AI industry has become a major concern for governments, utilities, and tech companies. A new study indicates that the upcoming generation of artificial intelligence could exacerbate this issue significantly.

      Researchers from the Korea Advanced Institute of Science and Technology (KAIST) have released what they claim is the first thorough analysis of the energy costs associated with AI agents—AI systems that can reason, plan, and execute tasks independently. Their research reveals that these systems may consume as much as 136.5 times more energy per query than traditional generative AI models, prompting doubts about whether the infrastructure supporting future AI can meet the anticipated demands.

      Enhanced AI brings a considerable increase in electricity expenses

      Unlike standard chatbots that produce a single response to a prompt, AI agents repeatedly access large language models (LLMs), search the internet, run code, use calculators, and interact with external applications to manage complex tasks. While these features significantly enhance their utility for research, programming, and automation in workplaces, they also demand much greater computational resources.

      Led by Professor Minsoo Rhu from KAIST’s School of Electrical Engineering, the research team classified AI agents as a new type of data center workload and evaluated their computational needs in practical scenarios.

      The results were notable. The researchers discovered that AI agents could increase response latency by up to 153.7 times compared to traditional rationale-based approaches. Even more surprising, the costly GPUs that support these workloads were idle for as much as 54.5 percent of the execution time while awaiting external tools to complete their tasks. In essence, the hardware continues to draw power even when it is not directly engaged in AI computations.

      Energy consumption also scaled dramatically. Operating an AI agent with a 70-billion-parameter language model, similar in size to contemporary commercial AI systems, averaged 348.41 watt-hours per query. This figure is approximately 136.5 times the energy required by a standard chatbot for a simple question.

      To grasp the wider implications, the team simulated a future in which AI agents manage 13.7 billion requests daily—akin to Google’s daily search volume. In this scenario, AI infrastructure would need roughly 198.9 gigawatts of electricity, nearly half of the average energy consumption across the entire United States and far exceeding the capacity of current AI data centers.

      The hidden expenses of intelligence

      These findings come as companies like OpenAI, Google, Microsoft, and Anthropic invest heavily in agentic AI, promoting it as the next significant advancement beyond conversational chatbots. However, the study asserts that merely enhancing AI models is no longer sufficient. Future advancements will rely equally on developing more efficient semiconductors, optimizing GPU usage, designing smarter data centers, and expanding power infrastructure.

      Professor Rhu points out that the research indicates a shift in competitiveness from creating “smarter AI” to developing more energy-efficient AI. The team believes that future AI progress will necessitate a co-design strategy, where models, AI chips, servers, and energy systems are optimized together to keep operational costs manageable and ensure sustainable AI growth.

      The paper, titled “The Cost of Dynamic Reasoning: Demystifying AI Agents and Test-Time Scaling from an AI Infrastructure Perspective,” was presented at the IEEE International Symposium on High-Performance Computer Architecture (HPCA) earlier this year. The researchers have also open-sourced their AI agent benchmarks, aiming to promote further research on reducing one of AI’s fastest-growing—and frequently overlooked—expenses: electricity.