I could see arguments be made in both ways here. If GPUs end up being more efficient/powerful (like today) it could induce even more demand, but also if CPU gets within ~20% of how fast you can do something with a GPU, people might start opting for something like Macs with unified memory instead of GPUs.
Today a CPU setup is still nowhere near as fast as a GPU setup (for ML/AI), but who knows how it looks like in the future.
> it is unclear that throwing more compute actually expands what is possible
Wasn't that demonstrated to be true already in the GPT1/2 days? AFAIK, LLMs became a thing very much because OpenAI "discovered" that "throwing more compute (and training data) at the problem/solution expands what is possible"
Today a CPU setup is still nowhere near as fast as a GPU setup (for ML/AI), but who knows how it looks like in the future.
> it is unclear that throwing more compute actually expands what is possible
Wasn't that demonstrated to be true already in the GPT1/2 days? AFAIK, LLMs became a thing very much because OpenAI "discovered" that "throwing more compute (and training data) at the problem/solution expands what is possible"