@paul You could theoretically have said the same about thunderbolt eGPU’s on M1 machines. Marcan had posted back on that old social network a couple years ago that the actual problem was related to the base addresses available on the PCIe lanes for hardware and at least in Linux's case too much stuff was hard-coded to rely on how it works elsewhere. (Warning: This is my understanding of what he said)

I had asked a few months ago if that changed for M2 and he had not looked into it yet.