That paper is more about training compute costs and emissions than about inference. Unfortunately, lots of published work in the training emissions area is rife with misinformation. This paper I and others wrote has more data on this topic:

→ View original post on X — @jeffdean