eleitl (Discord): We have a cluster with 5 peers, more than 20M blocks in their repositories (each), about 70.000 pins distributed among the 5 peers, using more than 2TB space on each.

Machines are using about 5GB of RAM each,  for go-ipfs and ipfs-cluster-service. (go-ipfs is in a more recent commit than last stable release). I remember back in the day when machines would be approaching 40GB of RAM.

Regarding CPU, they have 12 cores each, and the are at 300%-400% CPU normally (so about 4 cores), specially due to the re-providing process for all the keys.

go-ipfs is configured using the Badger datastore (should handle millions of keys better).

Hope that's useful!