custom_json ops don't affect the ram usage of consensus. These ops are simply validated and placed into the block log -- stored on disk. It's up to any second layer solution to determine what (if anything) to do with the data; the witnesses have only validated the account truly signed it and at what time, nothing about the content.
Depending on how the second layer (e.g. Steem engine) interprets the data, it could affect the ram usage of an operator in that layer. So they are the ones that must be careful with their design.
RE: Want Larger Custom JSON Payloads In Your Steem Operations?