At some point, we'll max out serial again and have to gang them together into parallel.
You mean PCI-express x4 x8 and x16 ? The old style parallel communication where all parallel paths have to be in sync with one clock are out and I don't think they will ever come back.
Going to discrete components will only be able to be done if the I/O bottleneck in some way can be solved. Currently the internal speed of a CPU is too big compared to bandwidth and latency of external IO making putting everything on one chip cost effective. Currently no solutions are in sight for this fact.
greets,
Staf.