If you're going to pull this, you need to know in advance that it's necessary, that's not some minor thing that everyone is going to just do automatically, it's an extra, complicated step it's easy to screw up.
That said, I acknowledge the possibility of compromise in the aforementioned scenario, but once again I don't understand people who always jump to the "I am vulnerable to this narrow threat model, ergo, I should not bother to protect myself against any threat models". Especially when the measures you can take to protect yourself most likely will address the actual threat models you are liable to encounter.
mid to high range ones you can though so that might be a workable solution there.
Also, they can still rip the ram out and nitrogen it.
With Virtual Private Servers, it's trivial to do the virtualized version of this through the hypervisor.
At this point you have pretty much everything on hand to help you fool the customer into believing everything is ok (eg: replace the whole physical machine with a vm...).
Come to think of it -- if you know the hardware in the machine, you probably could replace the machine with a vm anyway -- just mirror the disks and go.