Autonomous AI agents consume up to 136 times more energy than chatbots

The study that changes the energy conversation

When you interact with ChatGPT to ask a question, the energy consumption is significant but manageable. When that same model acts as an autonomous agent — planning, using tools, executing code, browsing the web — consumption can multiply up to 136.5 times. That's the main finding of a study published by researchers from the Korea Advanced Institute of Science and Technology (KAIST) at the 32nd IEEE International Symposium on High-Performance Computer Architecture, presented in February 2026. The paper, which has received massive attention in media outlets like Gizmodo, offers concrete data that is redefining the conversation about AI's energy cost.

The problem is not just peak consumption: it's what happens while the agent waits. This study comes at a time when the AI industry is heavily betting on autonomous agents, from code assistants to enterprise automation systems. The additional energy consumption could have significant implications for data center operating costs and AI's carbon footprint. Moreover, the finding that GPUs are idle more than half the time suggests there is enormous room for improvement in current system efficiency.

What did the KAIST team measure?

Professor Minsoo Rhu's team designed the first systematic analysis of AI agents' energy consumption as a new workload category for data centers. They measured real end-to-end energy consumption, from when the user prompt arrives to when the final response is delivered. The most striking results:

• An AI agent consumes up to 136.5 times more energy per query than a conventional generative model operating as a chatbot.
• Response latency can be 153.7 times higher for complex agentic tasks.
• GPUs can be idle up to 54.5% of the time during execution of an agentic task, waiting for responses from external tools (APIs, browsers, code executors).
• An agent based on a 70-billion parameter model consumes an average of 348.41 watt-hours per query.

For perspective, a typical chatbot query consumes about 2.55 watt-hours, according to study estimates. An agent performing multiple reasoning steps, calling APIs, and processing results can easily spike that number. The researchers also observed that consumption varies greatly by task: from 10 times more for simple tasks to 136 times for complex scenarios like generating reports with real-time data.

Historical context: the evolution of energy consumption in AI

This study does not come out of nowhere. Since 2020, several works have warned about the growing energy consumption of large language models. For example, a 2023 study from the University of California, Riverside estimated that training GPT-3 consumed about 1,287 MWh, equivalent to the emissions of 120 cars for a year. However, the focus now shifts from training to inference, and more specifically to agentic inference. Unlike chatbots, which generate responses in a single step, agents execute reasoning-action-observation loops, multiplying iterations and thus consumption. This phenomenon resembles what happened with cloud computing: initially the cost of idle instances was underestimated, and later solutions like spot instances emerged. Similarly, the 54.5% GPU idle time in agentic tasks represents waste that the industry must address.

Impact on businesses and users

For companies deploying AI agents, this means energy costs could be much higher than anticipated. An agent processing 1,000 queries per day could consume up to 348 kWh daily, which at an average industrial price of $0.10/kWh amounts to $34.8 per day, or over $10,000 per year per agent. In a data center with thousands of agents, the impact is massive. Infrastructure providers like Nvidia, AMD, or Google may need to redesign their architectures to handle agentic workloads more efficiently. For instance, Nvidia has already begun optimizing its GPUs for inference, but idle time during external tool waits suggests new techniques are needed, such as asynchronous execution or GPU virtualization. It also opens the door for new startups to optimize GPU usage during idle times, similar to what Docker did with containerization. For end users, although they don't directly pay the energy bill, the cost could be passed on through higher subscription prices or reduced availability of free services.

Market consequences

The KAIST study has direct implications for the AI market. Companies like OpenAI, Google, and Anthropic, which are betting on autonomous agents, will need to reconsider their business models. For example, OpenAI recently launched 'Operator', an agent that can perform tasks on the web, and this study suggests its operating cost is much higher than ChatGPT's. This could justify higher prices or tiered subscription models. Additionally, hardware manufacturers like Nvidia could see an opportunity in designing specialized chips for agentic workloads, with better latency management and standby consumption. AMD and AI chip startups could also benefit if they offer more efficient solutions. On the other hand, virtualization and orchestration software companies, like VMware or Kubernetes, could develop tools to minimize GPU idle time, for instance through parallel task scheduling or dynamic resource reallocation.

What should readers know?

Not all AI agents will consume 136 times more energy; the factor depends on task complexity and the number of tools used. However, the KAIST study is a wake-up call: energy efficiency must be a priority in designing agentic systems. Users and businesses should consider the environmental and economic impact when adopting these technologies. Moreover, the researchers note that current efficiency metrics (like cost per token) do not capture the real cost of agents, and propose new metrics such as 'energy cost per completed task'. In summary, we are facing a paradigm shift: from conversational AI to agentic AI, and with it, a new set of energy challenges that the industry must solve for the promise of autonomous agents to be sustainable in the long term.