if all those companies go bankrupt, you could also buy their hardware on the cheap.

its almost guaranteed imo that as the model quality evens out, inference cost will drop towards 0 at insane speeds, given how well llama works as an asic.