The race to continue training and researching, however, is drive by competition that will fall away if competitors also can't raise more money to subsidise it.
At that point the market may consolidate and progress slow, but not all providers will disappear - there are enough good models that can be hosted and served profitably indefinitely.
For some uses, sure. But for plenty of uses that can be provided in context, RAG, or via tool use, or doesn't matter.
Even for the uses where it does matter, unless providers get squeezed down to zero margin, it's not that new models will never happen, but that the speed at which they can afford to produce large new models will slow.
At that point the market may consolidate and progress slow, but not all providers will disappear - there are enough good models that can be hosted and served profitably indefinitely.