> "AI" (and don't get me wrong I use these LLM systems constantly) is off the charts compared to normal data centre use for ads serving.

It wasn't only about serving those ads though, traditional machine-learning (just not LLMs) has always been computationally expensive and was and is used extensively to optimize ads for higher margins, not for some greater good.

Obviously, back then and still today, nobody is being wasteful because they want to. If you go to OpenAI today and offer them a way to cut their compute usage in half, they'll praise you and give you a very large bonus for the same reason it was recognized & incentivized at Google: it also cuts the costs.