Is it because they will have to communicate back errors during training? I forgot that training these models is more of a global task than proteins folding. In that sense this is less parallelizable over the internet.