One thing that isn't clear from the website is that it requires two GPUs, one for the host and one for the VM.

Wouldn't be it possible to let Linux/Xorg give up GPU for a time Windows is used? Which could still let me access it via VNC from Windows. My native platform is Xubuntu on AMD Renoir laptop.

If this uses vfio then no, the host never sees the GPU, it's blind to it from boot.

What is stopping GPU hot plugging from working?

There was this recent improvement at least: https://phoronix.com/scan.php?page=news_item&px=Linux-5.14-A...

edit: seems there are some howtos - eg https://github.com/joeknock90/Single-GPU-Passthrough