The advantage of relying on a centralized relay caching proxy is that some things are faster but the downside is you’ve got a single point of failure for all users of the decentralized app. 

I think we need to do work to reduce the need of these centralized caching servers. Or make them more horizontally scalable as part of the protocol instead of a custom overlay.  

 Deploying many horizontally scalable instances is the way. 

One caching service is not interesting. Dozens (or hundreds) of instances is very interesting and could help scale nostr imho.