为什么 AI Agent 的耗电量远高于单次提示

When we talk about the energy cost of AI, it is no longer useful to think only about one chatbot prompt at a time. The move from single-shot prompts to agentic workflows changes the math in a very real way. That is the core point climate scientist Zeke Hausfather was making in a blog post flagged by Wired’s Molly Taft. Instead of estimating the electricity behind one response, he tried to estimate what happens when an AI assistant keeps working through a task, step by step, over and over again. From One Prompt to an Ongoing Workflow A normal chat interaction is easy to picture: you ask a question, the model answers, and the exchange ends. An AI agent is different. It may run a command, read a file, call a tool, then continue with the next step. Each of those actions expands the amount of text the model has to process. That is where the energy profile starts to change. Hausfather’s approach was to measure the number of tokens his AI assistant consumed and use that as a proxy for energy demand. Based on that accounting, he estimated that his average daily use landed somewhere in the range of 1.2 to 5.9 kWh. The exact figure depends on how much the agent does, but the takeaway is that the load is not trivial. Why Agentic Systems Multiply the Work The mechanism is straightforward once you see it. As Hausfather put it, every time an agent takes a step, such as running a command, reading a file, or calling a tool, the model re-processes its entire accumulated context. That means the model is not just handling the latest line of input. It is repeatedly working through the full history of the task. That repeated re-processing is what makes agentic usage different from a one-off prompt. The context keeps growing, and the system keeps paying the computational cost of carrying that context forward. Hausfather also noted another useful detail: the text a user actually sees, meaning the model’s output, is only around 0.4 percent of the total tokens processed. In other words, most of the computational effort is invisible from the outside. For developers, that is an important mental model shift. The output may look small and focused, but the underlying work can be much larger because the agent keeps re-evaluating everything it has already accumulated. What This Means at Population Scale A single user’s daily consumption may not sound alarming on its own. But the story changes when you think about many users running many agentic workflows all day long. That is why the article’s climate concern is not about one isolated interaction. It is about aggregate demand. If lots of people are using assistants that repeatedly re-process growing context windows, the energy use adds up quickly. This is also why simple comparisons to a single chatbot prompt can be misleading. The old framing misses the multiplier effect that comes from iterative tool use, long context, and repeated internal processing. The Part We Still Do Not Know Even with Hausfather’s estimate, the full climate impact remains hazy. The source does not claim a precise global number, and it does not pretend that the exact toll is already settled. That uncertainty matters. It means the best current read is directional rather than definitive: agentic AI can use substantially more electricity than the casual “ask a question, get an answer” model people often imagine, but the exact climate cost still depends on how these systems are used in practice. Why Builders Should Care If you are building with agents, the practical lesson is not that you should avoid them entirely. It is that you should understand where the cost comes from. The expensive part is not just the visible response. It is the repeated processing of growing context as the agent moves through commands, files, and tools. That has implications for how you design workflows, how often you let the agent loop, and how much history you keep attached to each step. The source article does not offer a recipe for reducing consumption, but it does give a clear engineering warning: agent behavior scales differently from simple prompting. If you are evaluating the footprint of an AI feature, you need to look at the whole workflow, not just the final answer. That is the real shift here. The energy story is no longer about one prompt in isolation. It is about the cumulative cost of a system that keeps thinking, checking, and re-processing as it acts.