Great initiative!
I'm planning (as already mentioned in my last post) to provide some benchmark results for different node configuration to help make better decisions regarding infrastructure needed to run various Steem related services (so it would be for both consensus and API nodes and later whole infrastructure supporting it)

Server(s) costs itself even over-scaled and powerful enough to handle all expected load are not that big compared to infrastructure features such as HA, (redundancy, load balancing, self healing) etc etc and of course human labor.

Currently I would suggest using smaller servers with specialized steemds (consensus node for handling subset of calls, then account history (ops can be whitelisted differently for different nodes and different use cases) etc etc) Don't throw requests than can be handled by consensus nodes at your full nodes.