AI Agents Are Using an Outrageous Amount of Electricity

In today's age of AI agents, it's crucial to recognize that calculating their total electricity usage has become a significant challenge. Unlike standalone chatbots, these sophisticated models are composed of various tools and application layers, requiring a vast amount of power to operate autonomously. Private AI labs are largely silent on their internal energy consumption, making it difficult to determine the full impact of our AI assistants on the electrical grid.

Climate scientist Zeke Hausfather, via a blog post highlighted by Wired's Molly Taft, estimated the energy demand of his daily sessions with Claude Code, an AI-coding assistant from Anthropic. By analyzing the number of tokens his AI assistant consumed, Hausfather determined that his daily usage ranged from 1.2 to 5.9 kWh, which is more electricity than two refrigerators running continuously.

The majority of token consumption, about 3.2 billion over eight weeks, was due to the agent "re-reading its own working memory" at each step, with only 0.4 percent of tokens being new information. Over a year, this energy usage would be equivalent to running an electric clothes dryer or driving an electric car around 11,500 miles in California.

While this might seem insignificant, it contributes to a growing source of emissions, worsening a global issue where temperatures are rising and emission reduction goals are falling behind. The true climate impact remains uncertain, as AI sustainability researcher Boris Gamazaychikov's upcoming research could revise our understanding of the true cost of AI agents.