NVIDIA is reportedly planning an upgrade for NVLink interface to accommodate higher bandwidth for IBM servers.
NVIDIA NVLink 2.0 and Volta
IBM server equipped with Power9 CPU and NVLink 2.0 interface would support up to six Volta-based comput co-processors. NVLink 2.0 would benefit from faster interlink between the CPU and GPU (25 Gbps form 20 Gbps), which translates into 50 GB/s bi-directional bandwidth.
Meanwhile PCI-Express 4, which should also arrive next year, would offer a bandwidth up to 64 GB/s. As a comparison, a cluster of four Tesla P100 NVLink co-processors has a bandwidth of 160 GB/s (4* 40), which may explain why NVIDIA has not decided to replace NVLink with PCIe 4.0, and probably never will.
The presentation also confirmed that upcoming Tesla V100 (name is just a guess) would also be designed in SXM2 form factor (which basically confirms it has High-Bandwidth-Memory). Whether Big Volta (aka GV100) would be offered in two variants (computing and gaming-oriented), is yet unknown.