500 Nodes per server is on the higher side through we’ve seen this amount of instances work. It is possible with PMM 1.13 which uses Prometheus 2 which is much less resource intensive.
It is hard for me to give exact estimates but I would imagine pretty beefy box with no less than 16 cores and 32GB of memory. Amount of memory will be mostly driven by how many and what queries you run rather than data you’re streaming. For disk reasonably fast SSD should work.
If you have any load problems typically these are considerations to consider
Reduce resolution (ie METRICS_RESOLUTION) variable
Reduce amount of information you capture (for example table usage statistics or performance schema can generate a lot of data)
Let us know how it goes!