Depends entirely on whether or not Moore's law has hit a physical limit yet.
If we're still doubling computational every X years for the same cost, we'd be able to see same models cheaper. If we're however finally hitting physical limits, then yeah. It won't be cheaper and we can't just smash more context in it like there's no tomorrow.
Then the most likely breakthrough would have to be DNA storage (tri point storage, rather than bi point storage) and then we'd need fast read/write to DNA storage. As that'd theoretically allow for more context for the LLMs, context is the biggest bottleneck
But the fun thing about the scenario of open source/light weight models dominating. Is the fact all of the leading LLM companies would go belly up, it's the biggest bear take on Anthropic, OpenAi and the likes - They need to have people use the biggest, most expensive models they can't run themselves, that's their entire claim to trillion