Keeping old hardware running is very fragile. Most spare parts are decades old now too, and nobody knows if they still work properly.
The DMA might actually be the easy part.
The main concern is how many VM exits per second a modern CPU can handle. Even though the interface is "slow", you can probably do ~500k to 1 million IN/OUTs or MOVs per second to ISA. You'd have to trap (VM exit) every access to a IO port or memory within the ISA window.
During the pandemic I cobbled together a KVM-accelerated emulator for an old 386EX based board I had and trapping every memory access to IO was slower than the real hardware.
ISA motherboards still work and have been mass produced in the hundreds of millions. (I have about a dozen in my retro closet.)
The esoteric ISA expansion cards would be the bottleneck, not the motherboards.
With a modern 4 GHz CPU, that means your budget is only 4000 clocks per interrupt. The context save alone can make it hard to hit 1MHz, and you really want to have your interrupt code pinned to the L1 cache because a few memory reads will blow your entire budget.
As nice as USB4 is, I’ve never heard anyone claim it has single-digit microsecond latency, to say nothing of sub-microsecond latency. The typically-quoted ~20 usec latency of USB4 limits it to around 50 kHz for an emulated ISA card.
At the same time ISA card will never reach 40Gbps throughout with CPU clocking at tens of MHz only or often less.
https://www.reddit.com/r/vintagecomputing/comments/jbaw38/is... (2020)
ISA-over-USB, using real ISA cards in an emulator
https://en.wikipedia.org/wiki/Low_Pin_Count#ISA-compatible_o...