Check the sources, that does account for training.

The 3Wh figure is pretty much “pulled out of the ass”. Looking at other sources listed, it does seem true that similar-sized models use ~2-3Wh per query, amortized training included. So yeah I concede the point.