The host has more sources that contribute to the kernel entropy pool: network and block IO are more than any individual VM.

Keep in mind, though, that just accessing /dev/urandom in the guest won't need too much input from the host.

However, having a hwrng is obviously the preferred option where there is need for a lot of entropy.

Modern Intel processors have an RDRAND instruction which (cue conspiracy theorists) is a hardware RNG. https://en.wikipedia.org/wiki/RdRand

However, it's RDSEED that needs to be used as input if we use the Intel instructions.

https://software.intel.com/en-us/blogs/2012/11/17/the-differ...

Edit to add: we can obviously pass on the RDRAND instruction directly to the VM, but that's less portable, because not all servers in a server farm may have that instruction available, limiting live migration capabilities.

So we can just pass on output from RDSEED into the host's /dev/random, and feed the VM /dev/random by default as described.

other than the x86 instruction set there are some SOCs that embeds a hwrng directly, the one I used is Freescale iMX6 that has a FIPS-certified hwrng, which is very handy.