The first statement is not even wholly true. While training does take more, executing the model (called “inference”) takes much, much more power than non-AI search algorithms, or really any traditional computational algorithm besides bogosort.
Big Tech weren’t doing the best they possibly could transitioning to green energy, but they were making substantial progress before LLMs exploded on the scene because the value proposition was there: traditional algorithms were efficient enough that the PR gain from doing the green energy transition offset the cost.
Now Big Tech have for some reason decided that LLMs represent the biggest game of gambling ever. The first to find the breakthrough to AGI will win it all and completely take over all IT markets, so they need to consume as much as they can get away with to maximize the probability that that breakthrough happens by their engineers.
The thing is it’s been like that forever. Good products made by small- to medium-sized businesses have always attracted buyouts where the new owner basically converts the good reputation of the original into money through cutting corners, laying off critical workers, and other strategies that slowly (or quickly) make the product worse. Eventually the formerly good product gets bad enough there’s space in the market for an entrepreneur to introduce a new good product, and the cycle repeats.
I think what’s different now is, since this has gone on unabated for 70+ years, economic inequality means the people with good ideas for products can’t afford to become entrepreneurs anymore. The market openings are there, but the people that made everything so bad now have all the money. So the cycle is broken not by good products staying good, but by bad products having no replacements.