AI智能體的能耗是簡單聊天的600倍
氣候科學家zeke hausfather發現，使用claude code 8週消耗了約170 kwh電力——大概是標準聊天提示的600倍——揭示了自主ai系統隱藏的環境成本。
climate scientist zeke hausfather logged eight weeks of intensive use with claude code, tracking 1,138 inputs that generated over 14,000 model calls and processed 3.2 billion tokens. the total electricity consumption reached an estimated 170 kwh, averaging 150 watt-hours per prompt—a massive increase over the 0.24 to 0.34 wh figures often cited by google and openai for basic text queries. this disparity stems from agent architectures that repeatedly read cached context and generate millions of tokens per task, creating a usage pattern that more closely resembles running a clothes dryer than a quick search. as labs move toward autonomous agents running for days or weeks, this exponential rise in energy demand positions data center power supply and carbon emissions as major bottlenecks for ai deployment.