Datacenter-scale GPGPU compute is extremely efficient compared to what you would achieve with a network of independently operated distributed heterogeneous devices spread across the world. Just imagine the overhead for ensuring that a node doesn't poison the gradients. And that's just one very small part of the problem.
