Part of the problem with that approach is that you're still reliant on the host OS and have just pasted a new layer of abstraction over the old one. Kinda like the web. You haven't really fixed any of the problems of the underlying platform, while at the same time you've added complexity and bloat.
I think there is, perhaps, value in a middle ground. You could take the Linux kernel (chosen here for its existing driver support, openness, stable syscall ABI, and presence of virtualization and namespacing features), but throw away basically everything on top of it and build a new "operating environment". Then, over time, you could fork the kernel away, ditching legacy syscalls and features, maybe creating a stable driver ABI finally, etc.
I think there is, perhaps, value in a middle ground. You could take the Linux kernel (chosen here for its existing driver support, openness, stable syscall ABI, and presence of virtualization and namespacing features), but throw away basically everything on top of it and build a new "operating environment". Then, over time, you could fork the kernel away, ditching legacy syscalls and features, maybe creating a stable driver ABI finally, etc.