Computer adoption went from shared mainframes to personal computers then back to shared mainframes (cloud).

I think there's a good chance we see AI do something similar. It's mostly cloud now because OpenAI, Anthropic, etc all have the best hardware and best models that are a clear step above what you can run locally.
But we're going to hit a point where these models get good enough and cheap enough that it's diminishing returns to have a leading model. If you're a business owner it's probably better for you at that point to install a $5k server locally to run these things than it is to pay someone $5 every time you make a query.
Then eventually these things get so cheap that it makes sense to go back to the cloud. Why install a $5k server you have to maintain when the queries are now $0.01?
That middle section is the highest risk place for the AI trade to fall apart and I think it would be prudent to keep an eye on those building strategies to bridge from part 1 to part 3.