You're looking at today, look at tomorrow. Yesterday you needed a super-computing cluster to run frontier level LLMs. Today you need a $1600 tier video card. Tomorrow you're going to be running them on a plain old laptop. There's no way people can realistically priced out, even if LLM companies buy up literally everything - because state of the art today is tomorrow's second-hand fire-sale.
And today's models will be relatively obsolete in the future, yet they are already enough to do a lot with. I think the acceleration theory, that is - that the head company will gain an ever increasing edge due to recursive self improvement, is a scam to draw in naive investors. The far more likely scenario, and one that seems to match current trends way better, is that LLMs have an approximate peak of capability and we're going to asymptotically approach that peak.
Even if we're ever able to achieve true AI which I think is likely impossible for LLMs, there will likely be an asymptotic peak because at some point you can't just think your way to progress, but reach a point where you need more data (observation, experiment, etc) to keep moving forward.