As discussed, phones typically do not have GPUs and it is important to save battery. So many expensive computations need to be offloaded to the cloud to alleviate the cost. Of course, it is a matter of compromising different factors including the cost of executing a deep learning model on mobile, the cost of moving data off to the cloud, the cost of the battery for this transfer, and the cost of cloud computing. There is no single solution, and the optimal strategy depends on your particular situation.