Scientist Tracks AI Energy Use, Finds Agentic Prompts Consume More

Climate scientist Zeke Hausfather tracked his own artificial intelligence usage over eight weeks to estimate the energy consumption of AI models, particularly focusing on the increasing prevalence of agentic AI. Hausfather, who uses AI daily for research and data analysis, initiated this tracking project due to a lack of transparency from tech companies regarding the energy demands of their AI models. He meticulously logged 1,138 prompts entered into Claude Code, an AI model developed by Anthropic, to estimate the electricity consumed at data centers. This initiative stems from a broader concern about the environmental impact of AI, a field where technological advancements and usage patterns evolve rapidly.

Publicly available data on AI energy consumption is scarce. Google, for instance, estimated in the previous year that a standard text prompt on its Gemini model utilized approximately 0.24 watt-hours, a figure comparable to watching television for nine seconds. Similarly, OpenAI's CEO, Sam Altman, has stated that a typical ChatGPT query consumes around 0.34 watt-hours. However, Hausfather hypothesized that these figures, which likely reflect earlier, simpler AI interactions, no longer accurately represent current AI usage patterns. He noted that "in the AI world, a year is an eternity," highlighting the rapid shift from basic chatbot conversations to more complex, agentic AI applications. Agentic AI involves instructing an AI agent to orchestrate multiple sub-agents to perform extensive processes, which he believes requires orders of magnitude more energy.

Hausfather's personal tracking revealed a significant difference in energy consumption between standard chatbot queries and his agentic AI prompts. He calculated that the median energy usage for his prompts was approximately 150 kilowatt-hours. This figure is estimated to be 600 times greater than the energy consumed by a single chatbot query. His estimation method involved analyzing token usage, which are the fundamental units of text processed by AI models. Hausfather reasoned that there is a strong correlation between token volume and energy consumption, as energy costs are a substantial component of running these large AI models. While acknowledging that this correlation is not perfectly precise, he used it as a foundational element for his calculations.

Despite the higher energy consumption of agentic AI, Hausfather's findings suggest that even his intensive AI use over the eight-week period resulted in an energy footprint comparable to running an electric dryer for a year. This comparison aims to contextualize the energy demand within a household appliance framework. However, he also stressed the importance of considering the aggregate impact of AI. As AI adoption grows across various sectors and applications, the cumulative energy demand from data centers powering these operations is becoming a significant environmental consideration. The study underscores the need for more detailed and transparent reporting from AI developers on the energy efficiency of their models and the environmental implications of their widespread deployment.