Probably not practical on a scale useful to AI companies. Having the chips alone isn’t enough, you need boards design as server blades, with high bandwidth backplanes, power system, and other custom build things to make use of this. And if these are stripped parts, you would also need testing facilities and certification capabilities for thousands of chips. I don’t see how a striped 4090 would be better than the A800/H800. Lots of video cards are assembled in China, so I would assume they also have the plants to recycle them.

Not really, translation layer/virtual machine is standard in compute these days. KVM, VMWare, Rosetta on Apple, or Kubernetes/Dockers. RISC-V is building on RISC-4 instruction sets, so compilers and all will be there. These are barriers thought up by politicos with limited tech knowledge, then get surprised by the speed of change.