The bubble will pop, but probably not for a year or two.
From my understanding, a lot of the money riding in the grand AI circle of cash originates from OpenAI becoming 100s of billions of dollars profitable in ~1-2 years, which can only really happen if they create a general artificial intelligence, which, IMO just isn't going to happen. A bunch of companies have misestimated how long the GPUs will last, their final retail price, among a lot of other wacky things. Exactly when it will be is really anyone's guess, but I'd put money down on it being OpenAI.
Logic Learning Machines will still be around, it's just that they will get more specific. During my undergrad, I worked on a proposal for the Deep Underground Neutrino Expirement, and we were going to apply the fundamentals of Learning Models to act as a data filter out other particle detections rather than go through all the data ourselves (mind you the expirement itself is supposed to wind up with ~10^6 - 10^7 data lines, so it simply wouldn't be feesible to have us try to look at everything). Granted this is not a Large Language Model
My current job also is a mix of both, I work for a medical brain implant company to allow people to speak again. Same sort of thing, building it up on a general understanding of how someone thinks with a variety of test passages.
They also have the capacity to be useful in terms of making things more efficient. It's mostly a matter of test data. I think companies that have a specific use case for their AI and using specific information is actually a pretty good idea, and will be companies that last in the long run.
EDIT: The fall of truy Large Language Models trained on the entire interent will most likely bring down the cost of a lot of the things you mentioned, but the persistence of Logic Learning Machines and Medium Sized Language Models means there will still be an increase to pre-LLM prices. However, if other semiconductor companies can start to imitate TMSC's processes, or the foundries they setup in the states ever gets to the point where it can duplicate the process of 4? nm (can't remember the exact number) it would also bring down prices a decent bit.







