Climate scientist Zeke Hausfather has revealed that AI agents use significantly more energy than simple chat prompts, highlighting a critical gap in current energy consumption estimates. His analysis, based on detailed logging of his use of the programming agent Claude Code, shows that over eight weeks, 1,100 inputs generated 14,000 model calls and 3.2 billion tokens, consuming about 170 kilowatt-hours of electricity. This level of usage produces CO₂ emissions comparable to running an electric clothes dryer annually.
Hausfather’s data indicates that each prompt processed by the agent uses about 150 watt-hours, 600 times more than the median Gemini chat prompt reported by Google. His calculations show that a single day of agent use can exceed the electricity consumption of two refrigerators, with some intensive sessions using over 11 kilowatt-hours. These figures highlight the vast difference in energy consumption between simple chat interactions and complex agent-based tasks.
Hausfather’s analysis contrasts with the energy figures provided by Google and OpenAI, which downplay the actual energy use of AI systems. He notes that these estimates do not account for the extensive token generation, multimodal processing, and multi-agent systems involved in complex AI tasks. The study underscores the need for clean energy sources in data centers to mitigate the environmental impact of AI agents. Source: thedecoder
