For me just the convenience of having everything in one box. Simplifies networking too. I run home assistant, openwrt, OMV, an ubuntu dtop VM and a wordpress LXC on a little m93 I jacked up with 32Gb RAM. Backups are dead simple and it’s all on one little UPS.

Some might prefer metal for other reasons but simplicity and convenience are priorities for me, at least in my homelab.

Beyond the other reasons, the idea is that you can keep the hypervisor (host OS) minimal and stable and so it’s less likely to break. Whereas you can break the VM to your heart’s content and you’ll still have access to the host and don’t need to rely on live USB etc to recover (or even video access since you can still access the host over the network). You can also migrate the VM to a new system more simply since it’s just a VM with simpler, abstracted hardware that is less likely to run into driver issues.

One downside of a VM is that GPU passthrough/GVT-g isn’t always stable or performant (especially GVT-g), so you might want to keep some services on the host if they need GPU access (such as transcoding). You can still access the files in the VM over NFS/Samba from the host. Although I’d use containers (e.g. docker) at least, sticking with the principle of keeping the hypervisor OS as minimal as possible.

The rules now are generally: bare metal if that’s all the box will do, or it’s main task, container if it’s one of many services, vm if it’s a larger application you might migrate and i/o isn’t your limitation.

The line between container and vm is fuzzy, but bare metal means you’re making a design choice for that machine and if that or another application breaks the machine you’re screwed.

In a way freebsd is amazing for this, you put all applications in jails and don’t use the main userspace much, but the virtualiztion story isn’t quite there yet.