A striking statistic has been making the rounds this summer: AI agents use 136.5 times more electricity than a chatbot answering a single question. It sounds like the kind of number that should end the argument about whether agents are worth the energy. It comes from real research, which is why it has traveled so far. The trouble is what gets lost between the study and the headline.
The figure is genuine. It comes from a team at KAIST, South Korea's top engineering school, led by Minsoo Rhu with first author Jiin Kim, presented at the HPCA conference in February. The researchers ran the Reflexion agent framework on Meta's 70-billion-parameter Llama model and measured a single task drawing 348.41 watt-hours, against 2.55 watt-hours for a plain chatbot reply. Divide one by the other and you get 136.5. So far the viral claim is accurate.
What the number leaves out
The problem is that 136.5x is the top of the range, not the middle of it, as Forbes noted in unpacking the study. On the same model, a different agent framework called LATS came in at about 62 times the baseline. That is still a lot, but it is less than half the scary figure, and it shows how much the answer depends on which agent design you test. Picking the single highest result and presenting it as the typical cost is how a real measurement becomes a misleading one.
The deeper point is that the multiplier hides the actual issue. The KAIST team found that the GPUs running these agents sat idle up to 54.5 percent of the time, waiting on an external API to answer or a piece of code to finish running. The energy is not mostly going into brilliant reasoning. A large share of it is going into expensive hardware held open, powered on, doing nothing while it waits. One framework in the study ran 153 times slower than a baseline query, and that latency is where much of the cost piles up.
Why this matters more than the scare figure
Read that way, the finding is less a doomsday energy statistic and more an engineering to-do list. Idle time is fixable. Better scheduling, batching, and smarter handling of the waits between steps could cut the waste substantially, and that is closer to what the researchers were actually pointing at than "agents are 136 times worse for the planet."
None of this means agent energy use is trivial. Running a multi-step agent genuinely costs more than a single chatbot reply, sometimes a great deal more, and anyone budgeting compute for agents on the basis of chatbot-style usage will get an unpleasant surprise. That part of the concern is fair. What is not fair is treating the worst-case number from one framework as the fixed price of the technology.
The honest version is less quotable but more useful. Agents can be far hungrier than chatbots, the exact multiplier swings wildly with how they are built, and most of the gap is waste that better engineering can attack rather than an iron law of physics. It sits alongside the other energy numbers that get flattened in the retelling, like the claim that a single AI query drinks a bottle of water. The measurement is real. The certainty wrapped around it is not.
Commentarii · 0