Cryptographically random bit strings are more likely to be collision-free than whatever UUID scheme you can cook up. The reason is that finding a collision among (uniform, independent) random n-bit strings is extremely low as long as you produce fewer than roughly 2^(n/2) of them, whereas generating the same UUID according to a deterministic process is relatively plausible if you consider non-synchronized clocks and virtualization.
