I took a look at my own Claude Code use over the past 8 weeks to calculate agentic energy use.
-
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.
-
The biggest lever we have is to decarbonize electricity – and push AI companies to invest more heavily in clean energy like solar+storage, advanced nuclear, enhanced geothermal, etc. Using energy isn't inherently bad, its all about the impacts of producing that energy.
-
The biggest lever we have is to decarbonize electricity – and push AI companies to invest more heavily in clean energy like solar+storage, advanced nuclear, enhanced geothermal, etc. Using energy isn't inherently bad, its all about the impacts of producing that energy.
One potential silver lining: net-zero needs ~3x more clean electricity, and the main barriers are permitting, transmission, interconnection, not tech. If AI's capital and urgency fund firm clean power and grids, it could help clean the grid. Behind-the-meter gas, and it won't.
-
One potential silver lining: net-zero needs ~3x more clean electricity, and the main barriers are permitting, transmission, interconnection, not tech. If AI's capital and urgency fund firm clean power and grids, it could help clean the grid. Behind-the-meter gas, and it won't.
Check out the full analysis over at The Climate Brink: https://www.theclimatebrink.com/p/the-real-energy-use-of-agentic-ai
-
Check out the full analysis over at The Climate Brink: https://www.theclimatebrink.com/p/the-real-energy-use-of-agentic-ai
@hausfath This looks very interesting. I caught a typo in the "Credible estimates" link at the top of the article. Here is the URL from the article link:
https://eta-publications.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report)It works if I take off the trailing ")".
-
Check out the full analysis over at The Climate Brink: https://www.theclimatebrink.com/p/the-real-energy-use-of-agentic-ai
@hausfath The other trend is to local models. My M5 64GB laptop goes from a few watts to 50W or so (it has a 96W power supply) and the fans come on when it’s thinking hard. Thats still much less than sending prompts away to run on Claude.
-
One potential silver lining: net-zero needs ~3x more clean electricity, and the main barriers are permitting, transmission, interconnection, not tech. If AI's capital and urgency fund firm clean power and grids, it could help clean the grid. Behind-the-meter gas, and it won't.
@hausfath The other silver lining is that wind+solar+batteries is the lowest cost energy, so there’s an economic incentive to move away from other energy sources. Gas and other short term fixes for the capacity and distribution shortages should have a fairly short shelf life.
-
@hausfath How much on a clothes dryer? Our annual use is 31kWh!
-
Check out the full analysis over at The Climate Brink: https://www.theclimatebrink.com/p/the-real-energy-use-of-agentic-ai
@hausfath It's great to have numbers for the energy use of agentic AI. It makes sense to me that it is vastly higher than simple questions to a chatbot.
I also didn't realize the extent of efficiency improvements over the past decade. It's incredible. I'm glad you mentioned Jevon's paradox because it's definitely happening.
So efficiency is good, but it won't save us. Clean energy sources will help. It would be great if we (in the US) could tax or ban new fossil fuel energy sources.
-
@hausfath very interesting thanks.
For my understanding, where you talk about the reality being 150 to 1200 times higher, does that higher estimate include the energy used in model training? I don't think that's the point you're making but I'm just not clear!
-
B bettina@mastodon.nu shared this topic
