The use cases that this addresses are ones where you cannot train models in the same system/space/network as where the data exists.