One use case that recently came up in discussion is in a build farm. Lightweight containers with the possibility to use AUFS and RAM disks make for fast, isolated compiles, and they also seem great for automated testing. Having all of that managed with the same as more long-lived, traditional VMs is appealing. (Whether doing that in the same cluster as those longer-lived VMs is wise is a much more context-specific question.)

I can see similar arguments for lots other typical network services where a fully virtualized VM is unnecessary overhead (there's no need for full isolation or different OSes), but the logical isolation is valuable.

-j