Well that depends on your app/system and how much processing you do for each network interaction. Some of us don't really receive and parse 67 billions HTTP requests + json every second. If tail/mean latency increases a bit, well, let's concentrate HW to avoid network interactions, at least network reception/parsing.

Just looking for a way out of the mitigation quagmire for people/orgs who own the hardware and don't share it with external untrusted parties. For all the talk about threat modelling, everyone seems to think we all want -20% perf on the whole machine, while only one or two processes do net I/O...