The real problem I have is the low network performance. Yes, yes, before everyone jumps all over me and points to Jupiter etc.. I understand the problems in Pb/s bisection bandwidth for the large datacenters. That doesn't change the fact that I don't need an entire datacenter worth of stuff.. but I do need an Amdahl-balanced cluster. So big machines with wimpy (20Gb non-RDMA) networks prevent me doing my HPCish workloads on GCE.

Followed by waiting on GPUs and other user accessible accelerators of course.