Rows of servers inside a data center. AI agents draw far more power per task than chatbots, and much of the bill is paid while the hardware waits.
getty
The statistic making the rounds this week is that AI agents burn 136.5 times more electricity than a chatbot. It comes from a real study, the number is real, and it is close to the least useful thing in the paper it was pulled from. The more you understand what the researchers actually measured, the worse the picture gets for anyone budgeting compute over the next three years.
The work comes out of KAIST, South Korea’s top engineering school. The team, led by Minsoo Rhu with Jiin Kim as first author, measured the energy cost of AI agents against ordinary single-turn chatbot queries, and presented the results at HPCA, the main computer-architecture conference, in February. The headline figure is genuine. Running the Reflexion agent framework on Meta’s 70-billion-parameter Llama model, a single task drew 348.41 watt-hours, against 2.55 for a plain chatbot answer. That is the 136.5x everyone is quoting.
It is also the top of the range, not the middle. A different agent framework, LATS, came in at 62 times the baseline on the same model. The multiplier tracks how much reasoning you let the system do before it stops. Change the framework or the model size and the number moves by more than a factor of two. Quoting 136.5x as the cost of agentic AI is like quoting the fuel bill of the thirstiest car on the lot as the price of driving.
The Part Worth Reading Twice
Strip out the headline number and the mechanism underneath is the actual story. A chatbot takes your question, produces an answer, and stops. An agent loops. It plans, calls a tool, waits for a web API or a code run to return, reads the result, and plans again. Each pass back through the model costs energy, and the waiting costs more than people expect.
The KAIST team found that on tool-heavy tasks the expensive GPUs sat idle up to 54.5% of the time, drawing power while they waited for something outside the chip to respond. That is the finding a CFO should care about. The unit of cost stops being the query and becomes the completed task, and a large share of the bill is paid for doing nothing but holding hardware open. In a data center where the graphics processors are the single most expensive line, half-utilization is not an efficiency footnote. It is the economics of the product.
This is why the efficiency story that usually reassures everyone doesn’t fully apply. When a model gets cheaper to run per token, that helps the chatbot case directly. It does far less for the idle waits and the repeated passes that make an agent an agent. The same paper clocked one agent framework, on a smaller 8-billion-parameter model, running 153 times slower than a baseline query. You cannot optimize your way out of latency you created by design.
The Number That Isn’t A Forecast
Then there is the figure that traveled furthest and deserves it least. The study projects that if agent usage reached 13.7 billion requests a day, data centers would need roughly 199 gigawatts, about half the average power draw of the entire United States. It has been passed around as a warning about where the grid is heading.
It is not a forecast. The authors took their own small-sample, per-task energy numbers and scaled them to Google’s current daily search volume, a stand-in chosen because it is a large round number, not because anyone expects 13.7 billion agent tasks a day soon. It is a what-if built to show the shape of the risk, and the paper says so. Treating it as a demand projection misreads the source.
The genuine forecasts are more sober and, read carefully, harder to dismiss. Lawrence Berkeley National Laboratory put US data center use at 176 terawatt-hours in 2023, about 4.4% of national electricity, and sketched a path to as much as 12% by 2028. The International Energy Agency expects global data center consumption to more than double to around 945 terawatt-hours by 2030. Both are built on the volume of queries going up. Neither prices in the possibility that the average query itself gets an order of magnitude heavier as the industry shifts from answering to acting. That is the gap the KAIST work exposes, and it sits underneath every official projection rather than inside any of them.
What To Actually Do With This
The translation for companies is narrow and specific. If your AI roadmap for 2026 is agentic, and most enterprise roadmaps now are, then any compute budget priced on chatbot-style inference is quoting you the wrong number. Vendors selling agent products at a flat per-call price are carrying utilization risk that the KAIST idle figures make concrete, and that risk lands on their margins or eventually on yours.
Power procurement is the same problem one layer out. Utilities and hyperscalers siting the next wave of capacity are working from demand curves built on query growth. The autonomy multiplier is not in those curves. In markets where a grid connection already takes years to secure, underestimating load per task is not a rounding error, it is a stranded-capex risk.
The authors’ own conclusion is that smarter software alone won’t close the gap, and that chips, data centers and models need redesigning around the wait-and-loop pattern that agents impose. That is a longer job than a quarter or two. The near-term move for anyone allocating capital is duller and more useful. Ask the vendor what a completed agent task costs, not what a query costs, and ask what share of that cost is a GPU sitting idle. The honest answer will not be 136.5 times anything. It will be worse, because it will be specific to your own workload, and you will have to pay it every time.