Climatologist Zeke Hausfather tracked the usage of Claude Code over an eight-week period. According to data cited by The Decoder, 3,2illion tokens were processed during this time, resulting in approximately 170Wh of data center electricity consumption.
When calculated per request, energy costs were found to be roughly 600imes higher than those of a standard chat prompt. The Decoder notes that this result highlights the discrepancy between the low estimates published by Google and OpenAI and the actual load generated by more complex scenarios.
This conclusion is based on an observation from a single user presented in The Decoder's publication metadata; therefore, it cannot be considered an independent measurement of the entire industry. Comparative data from other users and companies, along with a unified calculation methodology, are needed for validation.
editorial commentary
Why it matters
A likely consequence is increased attention to the energy costs of long-duration AI usage scenarios and the methodologies for measuring them. The next observable signal will be comparable measurements from other users or developers. Significant uncertainty remains because only one user example and publication metadata are available.