Base on Arron Edell on MachineBox:
Traditionally, the problem with running machine learning on premise is that neural networks require a lot of GPU acceleration to run, especially when you’re doing training.
The benefit to the cloud is that you only pay for what you use, so in a world where training happens all at once, it makes sense to outsource that cost to a cloud vendor.