A lot of the research on this is dedicated to detecting so-called “blue pill” attacks, that is, a malicious hypervisor that is actively attempting to evade detection.
The classic trick to detect a VM is to populate the ITLB, run an instruction that must be virtualized (which necessarily clears out such processor state when it gives control to the hypervisor), then run some more code to detect if the ITLB is still populated. The first paper on it is located here, and a rather colorful explanation from a researcher’s blog and alternative Wayback Machine link to the blog article (images broken).
Bottom line from discussions on this is that there is always a way to detect a malicious hypervisor, and it’s much simpler to detect one that isn’t trying to hide.