With the cloud, there is really no limit for high-throughput calculations. We have scaled to 35,000 cores at most in test regime, and regularly use ~10,000 cores in production. For a single communication-intensive job (ie. large cell, hybrid functionals, GW) the low-latency interconnect options on some cloud providers allow to scale to ~16-32 nodes (see Fig. 2 from this recent manuscript). Alternatively, there are large memory nodes that we found very helpful for GW, for example. We presently administratively limit each cluster to 200 nodes (each with 16-36 cores) and allow users to use 10 nodes per single job.