@NewtonMark @benji_w @cstross @davidgerard

They're not really single purpose though, or at least not that narrowly single purpose. ML has existed before LLMs and there will still be useful ML models after the LLM hype wave passes and the cheap GPU capacity will probably lead to a lot of development of more focused and proven useful models. Sure it's expensive to run all the GPUs 24/7 but they don't need to, they can just have high burst capacity.