I suspect that another reason not all chipmakers are scrambling to make big investments in "AI" silicon capacity is that it's not yet proven that "AI" tools are all that valuable to end users.
So far, I've seen a whole lot of really fun toys. I can get neat ideas for my D&D campaign from ChatGPT, and I can remove objects I don't like from photos, and I can get nice drawings from descriptions (sort of, with lots of limitations, and the copyright implications are problematic). And lots more fun toys.
But where are the real world-killer applications? The ones that could remake the world, move hundreds of gigabucks, by automating away entire classes of jobs? AI customer support chatbots are nearly useless. Self-driving cars are not happening. AI-powered search is comically unreliable. Automatically generated news articles, legal documents, summaries, medical documents, whatever, they all must be carefully double-checked by a human that's qualified in the field. And we don't know how to fix these problems; we're just hoping really hard that they'll go away with a bigger model.
On top of that, it's not yet settled that scraping the Internet for training data is even legal, from a copyright perspective.
Do you really want to invest ten billions to build a new chip fab that will be operative five years from now, when it could well be that in a couple years the hype will die down and all that's left will be a few cool photoshop plugins and some games?