It all comes down to block size, and the fact that the network and software is not optimized to handle even full 32 MB blocks. This is the bottleneck. See stress tests if you don't believe me.