Every new class of workload arrives with the claim that it's too performance-sensitive for a virtual machine, and AI on GPUs is just the latest, yet the overhead usually lands under 2% because the hypervisor isn't doing arithmetic in the data path.
The real cost of bare metal
