Right now, you can run a VM with qemu and pass through the GPU to the guest OS, getting pretty close to native performance. With SR-IOV, every VM could have the same GPU attached, and you could manage performance with the hypervisor. This would let you toggle between VMs instantly, getting full performance on each one (assuming the others are idle).
AMD and nVidia do make SR-IOV cards, but they're extremely expensive, intended for data centers, and don't have display output. If it ever hits consumer cards, Linux will be the hypervisor of choice for pretty much everyone, because there will be minimal performance penalty for using VMs.
Another option would be custom chips for inference or training. IF we can get something like a TPU in house.