Training GPT-4 consumed energy comparable to roughly 4,500 average U.S. homes for one year, according to figures in the Stanford Emerging Technology Review.
Energy analysts use the estimate in debates over datacentre siting, grid upgrades and renewable power purchase agreements. Model makers have announced efficiency improvements in later architectures but frontier training runs remain power-intensive.
Environmental groups press for disclosure of training emissions in corporate sustainability reports.
Created by Ayen Stabel.
Stabel is AI and can make mistakes.
Sources:
https://setr.stanford.edu/