AI Agents Consume 136.5 Times More Energy Than Chatbots, KAIST Study Reveals Grid Strain Ahead

When It Comes to Energy Use, AI Agents Could Make Chatbots Look Like Pocket Calculators

Much has been made of the emergence of generative AI and its strain on the electrical grid due to the energy demand. So just wait until you see how much energy agentic AI consumes. A new research paper from the Korea Advanced Institute of Science and Technology set out to quantify the "hidden costs" of AI agents, and found they can consume up to 136.5 times more energy per query than generative AI models. There's a logic to the fact that AI agents require more processing power and energy than your standard generative AI query. Typically, LLM requests are a call-and-response: a person enters a query, and the model responds. But agentic AI typically requires multiple steps to execute a command. To do that, the researchers said, the agent must continuously ping its model to generate a new response as it reasons through all of the steps of its given task. As a result, there's a multiplier effect that takes place. According to the researchers, an AI agent running on a large language model of the scale of most commercially available AI models would consume an average of 348.41 watt-hours per query -- about the equivalent of keeping an LED light bulb on for a full day. That figure, they say, is about 136.5 times higher than the energy consumed by a generative AI query. The impact of agentic AI goes beyond energy consumption. The paper also examined response latency and found that agentic AI can take 153.7 times longer than a standard query. That matters because longer response times tie up computing resources. As an agent repeatedly pings a model to complete a task, the GPU can spend more time waiting than working. The researchers estimate that GPUs may sit idle for as much as 54.5% of the time while an agent executes a task, creating a level of inefficiency that is not present in more straightforward AI uses. Now, all of that would be one thing if agentic AI were just a concept being played with in a lab somewhere. The reality is that we're already getting inundated with AI agents. We have no real sense of just how many agents are out in the wild at this point. There are 200,000 verified agents registered on Moltbook, the social network for AI agents, and about 400,000 agents have reportedly been approved to use the stablecoin UDSC. Companies like Google have started to build agentic AI into the web browsing experience. It is already quite prevalent. Researchers also modeled a future in which AI agents generate 13.7 billion requests per day, roughly the same volume of queries Google Search currently handles. Without major gains in energy efficiency, they estimate that would create demand for about 198.9 gigawatts of power -- roughly half of the entire United States' current electricity consumption. I don't know if the planet can handle half of another U-S-A! But we're probably going to find out.

AI agents bring a steep energy cost beyond ordinary chatbots

A new study has found that AI agents can use up to 136.5 times more energy per query than an ordinary chatbot handling a similar request. The extra power comes not from smarter answers, but from how often these systems stop to think, reach for tools, and reconsider their next step. That gap turns a question of clever software into a question of electricity. If agents take on even a slice of everyday online tasks, the study projects power demands big enough to strain national grids and change how the infrastructure behind AI is built. Why agents burn power A conventional chatbot answers in a single pass. You ask, the model runs once, and a reply comes back. An agent works differently. It splits a goal into steps, calls outside tools such as a search engine or a calculator, reads what returns, and picks its next move, looping until it lands on an answer. The work behind the study, led by Professor Minsoo Rhu of the Korea Advanced Institute of Science and Technology (KAIST) School of Electrical Engineering, set out to measure what that looping costs. This was a direct measurement. Their setup ran five agent designs on tasks like question answering, shopping, math, and coding, while logging every call to the model. The first surprise was volume. A simple chatbot makes one call to its language model for each query, while the agents made far more, roughly nine times as many on average. One design pushed much further. A tree-search agent known as LATS averaged 71 calls to the model for a single request, opening separate calls to explore several lines of reasoning before settling on one. The idle GPU problem All that calling costs more than electricity. It costs time too. Each step waits on the one before it, since the model has to see a tool's result before choosing what to do next. That chain is hard to speed up. In the study's tests, an agent could take more than 150 times longer to finish than a chatbot answering a comparable question. The strangest cost hides in the hardware. The expensive GPUs - the specialized chips that run these models - sit idle while an agent waits for a web search or a calculator to answer. In some tasks, the chips stayed unused for as much as 54% of the run, powered on but doing no useful work. Power at grid scale Measured in watt-hours, the gap turns concrete. Running on a 70-billion-parameter model, close in scale to today's commercial AI services, a reflection-based agent called Reflexion averaged about 348 watt-hours to answer one query. The same model handling a plain chatbot request used roughly 2.5 watt-hours. Researchers had long suspected that agents were costly, but no one had measured the scale directly. Earlier work had shown that general-purpose generative AI is far hungrier than task-specific software, with one widely cited paper finding it can use orders of magnitude more energy per task. The new results extend that pattern into the looping world of agents. The projections scale that gap up to national grids. If agents someday handled 13.7 billion requests a day, matching current Google search traffic, the study estimates the power required could approach 200 gigawatts. That is close to half the average draw of the entire United States power grid. It also runs far beyond any AI data center now on the drawing board, including the multi-gigawatt sites tech companies are racing to build. Even at today's traffic, the demand remains high. A single ChatGPT-style query already uses about ten times the power of a web search, and separate research has warned that AI's overall appetite could soon rival that of a small country. The limits of scaling More computing does buy better answers. The catch is that the gains fade. Past a certain point, accuracy levels off while the electricity bill keeps climbing. The researchers call this approach test-time scaling - spending extra computation while a model works out an answer rather than making the model itself bigger. Those returns shrink fast. In one agent, wringing out a few extra points of accuracy near the ceiling cost about 31 times more computing than the same gain had cost earlier on. Not every route is a dead end. When the team let an agent explore several lines of reasoning at once, instead of one after another, accuracy rose and response time dropped as well. A smaller model using that approach came close to matching a much larger one while using less energy. The strain already shows up in tools people use. OpenAI's Deep Research feature, built for this kind of multi-step work, can take up to 30 minutes to answer a single request. To keep costs in check, the company limits how often subscribers can run it. Building smarter AI agents What was once murky is now measured. Until this study, the electricity behind an agent's reasoning was mostly guessed at; now there are hard numbers showing that the cost climbs with every extra loop, and that past a certain point the extra loops buy very little. Rhu sees the result as a change in what progress should mean. "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," said Rhu. The study argues for leaner designs rather than bigger ones. One option the authors highlight is mixing small and large models, letting a lightweight model handle routine steps and calling on a heavyweight only when a task truly needs it, an idea a recent paper argues could cut both cost and delay. For anyone using these tools, the trade-off is becoming visible in everyday limits, like caps on how many deep queries a subscription allows. The bigger question is whether the industry builds power plants to match brute-force agents, or builds smarter agents that need less power in the first place. The study is published in the 2026 IEEE International Symposium on High-Performance Computer Architecture. -- - Like what you read? Subscribe to our newsletter for engaging articles, exclusive content, and the latest updates. Check us out on EarthSnap, a free app brought to you by Eric Ralls and Earth.com.

AI's energy tax was already concerning. Research says AI agents are over hundred times worse

AI agents could consume 136 times more energy than today's AI, study finds The AI industry's soaring electricity demand has already become a growing concern for governments, utilities, and technology companies. But a new study suggests the next generation of artificial intelligence could make that problem significantly worse. Researchers from the Korea Advanced Institute of Science and Technology (KAIST) have published what they describe as the first comprehensive analysis of the energy cost of AI agents - AI systems capable of reasoning, planning, and completing tasks autonomously. Their findings show that these systems can consume up to 136.5 times as much energy per query as conventional generative AI models, raising fresh questions about whether the infrastructure supporting tomorrow's AI is ready for what's coming. Smarter AI comes with a much steeper electricity bill Unlike traditional chatbots that generate a single response to a prompt, AI agents repeatedly call large language models (LLMs), browse the web, execute code, use calculators, and interact with external software while solving complex tasks. While these capabilities make them significantly more useful for research, programming, and workplace automation, they also require far more computing resources. Recommended Videos Led by Professor Minsoo Rhu from KAIST's School of Electrical Engineering, the research team treated AI agents as a new category of data center workload. It measured their computational requirements in real-world scenarios. The results were striking. The researchers found that AI agents can increase response latency by up to 153.7 times compared to conventional chain-of-thought reasoning. More surprisingly, the expensive GPUs powering these workloads remained idle for up to 54.5 percent of execution time while waiting for external tools to finish their tasks. In other words, the hardware continues consuming power even when it isn't actively performing AI computation. Energy usage scales just as dramatically. Running an AI agent powered by a 70-billion-parameter language model, similar in size to today's commercial AI systems, required an average of 348.41 watt-hours per query. That's roughly 136.5 times higher than a conventional chatbot answering a straightforward question. To understand the broader implications, the team modelled a future where AI agents handle 13.7 billion requests per day - roughly equivalent to Google's daily search traffic. Under that scenario, AI infrastructure would require approximately 198.9 gigawatts of electricity, nearly half of the average power consumed across the entire United States and far beyond the capacity of today's AI data centers. The hidden cost of intelligence The findings arrive as companies including OpenAI, Google, Microsoft, Anthropic, and others increasingly invest in agentic AI, positioning it as the next major leap beyond conversational chatbots. But the study argues that improving AI models alone is no longer enough. Future progress will depend equally on more efficient semiconductors, better GPU utilization, smarter data-center design, and expanded power infrastructure. Professor Rhu says the research demonstrates that AI competitiveness is shifting from building "smarter AI" to building more efficient AI. The team believes future AI development will require a co-design approach, optimizing models, AI chips, servers, and energy systems together to keep operating costs manageable and ensure AI remains sustainable at 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) earlier this year. The researchers have also open-sourced their AI agent benchmarks, hoping to encourage further work on reducing one of AI's fastest-growing -- and often overlooked -- costs: electricity.

Advanced AI uses 136.5 times more electricity than standard chatbots, study warns - The Korea Times

As AI takes on more complex tasks, its hunger for electricity skyrockets, as visualized in this AI-generated image. Courtesy of KAIST The next generation of artificial intelligence could trigger a massive global energy crisis. A groundbreaking new study by the Korea Advanced Institute of Science and Technology (KAIST) has quantified the hidden electricity costs of advanced AI agents for the first time. The findings reveal that these autonomous systems consume up to 136.5 times more energy per query than the conventional generative AI tools we use today. While standard chatbots simply answer a question and stop, AI agents act like digital assistants. Given a goal like planning a vacation or managing a budget, they independently figure out how to do it. To finish the job, the agent will search the internet, make calculations and execute commands entirely on its own. However, this independence comes at a massive environmental cost. To complete a complex task, the agent has to talk to itself and rerun its core AI programming over and over again. This continuous looping means answers can take 153.7 times longer to generate. Even worse, expensive computer graphics chips sit completely idle more than half the time, burning electricity while simply waiting for external websites and apps to respond. The strain on global data centers is immense. The KAIST research team, led by Rhu Min-soo, found that a single complex request to an AI agent burns through an average of 348.41 watt-hours of electricity. If these agents become mainstream and handle an estimated 13.7 billion requests globally per day, the power grid will not be able to cope. Total data center electricity demand would skyrocket to roughly half of the average energy consumption of the entire United States. Rhu warned that tech companies cannot just focus on making software smarter if they want a sustainable future. Instead, the tech industry must completely redesign AI models, microchips, and data center power grids from the ground up to handle this massive new workload before the grid collapses. This article was published with the assistance of generative AI and edited by The Korea Times.