They don't call it out in the announcement, but this was published in the same week as the "Supercomputing" conference - GCP hasn't had an answer for HPC workloads thus far, so maybe this is their attempt.
It's not a proper answer for HPC without an actual low-latency low-contention RDMA fabric. Typical HPC systems have a large fraction of materials science computation that really needs it. There's an example of scaling cp2k (I think) on Azure, but I've not used Azure to be able to say how well it works compared with an actual HPC cluster where you can also monitor the fabric (which is quite important).
Fully agree, but the announcement is pretty bare-bones so no reason to think that adding InfiniBand is totally out of the question, particularly since this offering seems to be pretty far from a standard GCP experience.
Having experimented with IB on Azure, I can comfortably say that the problem is not so much the fabric but the general usability of Azure itself - hot garbage doesn't do it justice.