Good idea from the perspective of not exposing the code base. However, technologies such as remote Cuda/OpenCL which rely on remote execution of compute kernels

Good idea from the perspective of not exposing the code base. However, technologies such as remote Cuda/OpenCL which rely on remote execution of compute kernels in general require high-bandwidth and low-latency connectivity - this is especially true for deep learning / AI workloads, not necessarily for other applications which may have a higher computer to data transfer / synchronization ratios. The latency on a typical internet connection will likely stall the GPUs on a remote system, yielding little compute benefit.