In nearly every practical scenario, the scaling bottleneck lies in network throughput. Trying to scale by putting network I/O between computation and storage is a bit like trying to separate a rocket engine from its fuel tank with a garden hose.