This makes sense, but what happens when they stop burning cash on training runs and any of their competitors releases a better model that raises the ceiling?

They will have to train one that is comparable (or better), or the word will spread and users will move to the better model.