Before diving into the data, a quick clarification for anyone unfamiliar with the term: what is agentic AI? Until now, most of us used AI like a chatbot: we ask a question, it answers, and the interaction ends there. Agentic AI is a step beyond. Instead of just responding, the system acts on its own to complete a task from start to finish: it decides what steps to take, searches the internet for information if needed, executes code, corrects its own errors on the fly, and chains multiple actions together without a person having to guide each intermediate step. It's the difference between asking someone "how do I book a flight?" and telling them "book me a flight to Rome for next weekend" — and having that person handle everything themselves..
This capability is what's behind the new "agents" that are starting to be integrated into operating systems, browsers, programming tools, and enterprise applications. And while it sounds like a pure advantage, it comes with a cost that had barely been measured until now: the energy cost.
Until recently, talking about AI's energy consumption meant talking about chatbots: questions and answers, a relatively contained and already well-studied level of use. That framework no longer holds.
Until recently, talking about AI's energy consumption meant talking about chatbots: questions and answers, a relatively contained and already well-studied level of use. That framework no longer holds.
The problem is that, as adoption accelerates, the real cost of that autonomy had barely been rigorously measured. A study from KAIST (Korea Advanced Institute of Science and Technology), led by Professor Minsoo Rhu, was the first to quantify it systematically, comparing the energy consumption of real agents versus a conventional chatbot solving the same task. The result: up to 136.5 times more energy per query.
Why does an agent consume so much more?
The key isn't a single "heavy" step, but the multiplier effect of autonomy itself. A chatbot resolves a query with, generally, a single pass through the model. An agent, on the other hand, may need dozens of internal calls to complete the same task: it interprets the goal, generates a plan, invokes an external tool (a search engine, a code interpreter, an API), waits for the response, re-evaluates whether the result is sufficient, corrects course if not, and repeats the cycle as many times as necessary. Each of those loops runs the full model again.
And here a paradox emerges that the researchers themselves flag as one of the study's most significant findings: much of that added cost is not due to the GPU "thinking" harder, but to it waiting. While the agent waits for a response from an external tool, the GPU remains reserved and consuming energy — but idle. It's a structural inefficiency: you pay for an extremely expensive resource that, for a good part of the time, isn't doing any useful work.
The numbers, in detail:
An agent based on a 70-billion-parameter model (a scale comparable to current commercial services) consumes an average of 348.41 Wh per query. To put that in perspective, it's roughly what an LED bulb uses if left on for an entire day.
It's not just energy: latency also spikes. Response times can be up to 153.7 times longer.
And here's the most uncomfortable part: GPUs — that extremely expensive, high-demand hardware — sit idle for up to 54.5% of execution time, waiting for external tools to respond. You pay for computing power that, for much of the time, isn't being used.
The researchers went further and simulated what would happen if these agents handled a volume of traffic equivalent to Google's daily searches (13.7 billion queries per day). The required infrastructure would be around 199 gigawatts — close to half the average electricity consumption of the entire United States. It's a hypothetical scenario, not a near-term prediction, but it helps scale something that today goes largely unnoticed: data center expansion plans and electricity demand forecasts from the International Energy Agency are still largely calculated based on chatbot consumption patterns. If the adoption of autonomous agents scales at the pace the market anticipates, those forecasts could be falling short.
Why does this matter now?
Because agentic AI is not a distant promise: it's already being integrated into operating systems, browsers, office suites, and cloud platforms, and major tech companies are presenting it as the next productivity leap. The conversation about AI sustainability first focused on training large models; then on chatbot inference in production. Now a third chapter opens, that of autonomy, and it brings an energy bill of a different order of magnitude than the previous two.
For companies already deploying or evaluating agents in their operations, this has a very practical implication: it's not enough to budget for "queries" or "tokens," as has been done until now. You need to budget for completed tasks, with everything that entails: retries, tool calls, wait times, and reserved hardware even when idle. It's a very different way of measuring real cost than what we used with chatbots, and many organizations have not yet incorporated it into their calculations.
One important nuance: 136.5x is the peak observed in the study (using the Reflexion technique), not a universal average. Another configuration analyzed (LATS) yielded a figure of 62.1x. These are figures obtained in a specific benchmark (HotpotQA, with Llama-3.1 models), not a general law applicable to any agent on any task. But even as a range, the underlying message doesn't change: the gap between "responding" and "acting autonomously" is enormous.
The research team's own conclusion is clear: the challenge is no longer just designing more powerful models, but fundamentally rethinking hardware, model architecture, and data center design so that this new form of AI is energetically viable. Optimizing software is no longer enough.
In practice, this points to several lines of work we'll likely see develop over the coming years: smaller, more specialized models for the intermediate tasks of an agent (not every step needs the largest, most expensive model), orchestration systems that reduce the number of redundant calls, hardware capable of releasing or reassigning compute capacity while waiting for external responses, and data centers designed with intermittent workloads in mind, not the constant flows that chatbots generated. None of these solutions are trivial, and all require investment and R&D at a pace that today is not guaranteed.
The question we in the industry should be asking is not just "what can an agent do," but "at what energy cost does it do it, who absorbs that bill, and who will be able to afford it at scale." It's a business question, an infrastructure question, and — increasingly — an energy policy question.
Sources:
Kim, J., Shin, B., Chung, J., Rhu, M. — "The Cost of Dynamic Reasoning: Demystifying AI Agents and Test-Time Scaling from an AI Infrastructure Perspective", KAIST (arXiv), July 2026.
The Korea Times — "Advanced AI uses 136.5 times more electricity than standard chatbots, study warns" (Jul. 5, 2026)
Gizmodo — "When It Comes to Energy Use, AI Agents Could Make Chatbots Look Like Pocket Calculators"
Digital Trends — "AI's energy tax was already concerning. Research says AI agents are over hundred times worse"
Forbes — "AI Agents Don't Use 136 Times More Power. The Truth Costs More" (contextualizes and nuances the viral figure)