Currently, if you want to resell computing resources, you need to rent or buy a dedicated server, and run a hypervisor on it.
Containers enable a new class of reselling computing resources. Because you can run a container within a VM, you can resell computing capacity on a VM.
I think we are going to see another abstraction on top of "the cloud," due to this additional layer of reselling (new russian doll on the inside, new doll on the outside).
The physical abstraction is:
Datacenter > Floor Space > Server Rack > Server
The virtual abstraction is:
Server > VM > Container > Kubernetes|MESOS|...
Virtual is a 1:1 inverse of physical. Next step is datacenter interconnections (i.e. multihost kubernetes or whatever flavor of the month IaaS software people use).
What I'd rather see is an allocation layer for physical resources that just cordons off the whole machine (physical or virtual) by tenant as soon as previous tenant resources have been fully consumed, then reclaims hosts after usage subsides. So as a provider I still only have one cluster to manage, but as a consumer I still don't worry about another layer of abstraction slowing things down or pre-allocating resources.