Theoretical dedicated gpu vm that paravirtualizes to other guest VMs

There’s a different approach that works with newer CPUs: