I think the comparisons to images are a bit misleading, where the author implies that generating images are somehow more energy-efficient than taking the photos

I think the comparisons to images are a bit misleading, where the author implies that generating images are somehow more energy-efficient than taking the photos yourself. It basically ignores the fact that the very data these large models use came from real photos taken from real cameras etc. and that used real energy, which in the end should be included in the total training cost, i.e., in addition to the electricity consumed by the GPUs. This would have an impact in how the total cost can be amortised with usage. Of course the impact might be just to slow down a little the amortisation with use, but it still should be included for a fair comparison. Same applies to some of the other examples they give.