The energy consumption of AI has already raised alarms. Studies indicate that AI agents are over a hundred times more energy-intensive.

      The surging electricity demand in the AI industry has become an increasing worry for governments, utility providers, and tech companies. A recent 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 term the first thorough examination of the energy expenses related to AI agents—AI systems that can reason, plan, and perform tasks independently. Their research reveals that these systems can use up to 136.5 times more energy per query than traditional generative AI models, prompting new concerns about whether the infrastructure intended for future AI is prepared for the upcoming challenges.

      More advanced AI leads to much higher electricity costs

      Unlike standard chatbots that provide a single answer to a question, AI agents continuously engage large language models (LLMs), search the web, execute code, utilize calculators, and interface with external software while completing complex tasks. Although these features enhance their utility in research, programming, and workplace automation, they also demand considerably more computing resources.

      Guided by Professor Minsoo Rhu from KAIST's School of Electrical Engineering, the research team categorized AI agents as a distinct type of data center workload and assessed their computational needs in real-world contexts.

      The findings were remarkable. The researchers discovered that AI agents can prolong response times by as much as 153.7 times when compared to conventional chain-of-thought reasoning. Even more surprisingly, the costly GPUs that support these tasks remained inactive for up to 54.5 percent of the operation time as they waited for external tools to complete their tasks. This means that the hardware consumes electricity even when it is not actively performing AI computations.

      Energy consumption escalates just as drastically. Operating an AI agent powered by a 70-billion-parameter language model, which is similar in scale to current commercial AI systems, averaged 348.41 watt-hours per query. This figure is approximately 136.5 times greater than the energy used by a traditional chatbot responding to a simple inquiry.

      To assess the wider consequences, the team modeled a future where AI agents process 13.7 billion requests each day—about the same level as Google’s daily search volume. In this scenario, the power requirements for AI infrastructure would reach around 198.9 gigawatts, nearly half of the average electricity usage across the entire United States and significantly exceeding the current capabilities of AI data centers.

      The overlooked expense of intelligence

      These findings emerge as companies such as OpenAI, Google, Microsoft, Anthropic, and others are increasingly pouring investments into agentic AI, regarding it as the next significant advancement beyond conversational chatbots. However, the study contends that merely enhancing AI models is no longer sufficient. Future advancements will also rely on more efficient semiconductors, improved GPU utilization, smarter data center design, and a more robust power infrastructure.

      Professor Rhu points out that the research indicates a shift in AI competitiveness from creating “smarter AI” to developing more efficient AI systems. The team believes that upcoming AI advancements will necessitate a co-design strategy, optimizing models, AI chips, servers, and energy systems simultaneously to keep operating costs under control and ensure AI remains sustainable on a large scale.

      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) this year. The researchers have also made their AI agent benchmarks available as open-source, with the hope of promoting further efforts to reduce one of AI’s rapidly growing—and frequently overlooked—expenses: electricity.