A certain LLM booster wrote recently that models got cheaper because they're more efficient but proceeded to only talk about it in terms of *cost* and didn't mention power consumption at all.
Perhaps you failed to consider the hyperscalers are in a race to the bottom?