The Channel I/O worked on many devices and the OS. Far as cost, using AWS might not be cheap but I was thinking redundant VM's or dedi's with cost-effective hosts. That doesnt cost hundreds of thousands to millions a year.
Perhaps I'm missing something. I'm still unsure if you're saying that Channel I/O (or its equivalent) is missing from current x86 server systems. Does virtualization affect this situation?
Initially, you mentioned high utilization ratio and throughput. I think the former is a red herring [1], but I'm curious about the latter. I've certainly witnessed poorer I/O throughput under virtualization, but on bare metal, throughput doesn't seem to be limited (beyond the capabilities of the bus).
[1] e.g. it doesn't matter if CPU is pegged but I/O channels are at 10% if the workload is CPU-bound and CPUs are the expensive part to scale. Or substitute memory for CPU.
For example, it seems we already achieved that with disks, once embedded controllers attached to DMA-capabled HBAs became the norm.
A similar thing seems to have happened with NICs, as well as the ability to offload higher-level protocol processing (another mainframe-like feature).