2026-08-05 16:46 UTC
Published estimates of AI energy use per prompt (~0.3 watt-hours) appear reassuring small. But they are increasingly disconnected from how we actually use AI. Modern agentic tools use around 600 times more energy per prompt, as I discuss in my latest over at The Climate Brink https://www.theclimatebrink.com/p/the-real-energy-use-of-agentic-ai
Replies (2)
-
@hausfath@fediscience.org 2026-08-05 16:46
I took a look at my own Claude Code use over the past 8 weeks to calculate agentic energy use. Using three different methods I find that the typical agentic prompt uses 180 watt-hours (45 to 360 Wh). This adds up to around 3 kWh per day, around the energy of running two fridges.
-
@guigsy@mstdn.social 2026-08-06 10:20
@hausfath@fediscience.org the hardware is getting more efficient. But LLMs themselves are only seeing minor gains. To get marginal increases in quality of response, we're needing exponentially more compute power. This analysis only estimates direct energy usage. I've read that if you factor in the energy used for training, and pro rata it to users, it roughly doubles the energy used.