As for the 500ms I wasn't thinking only network overhead, but also things like complex sql queries with lots of joins etc. And maybe on top of that going through a json-rest/soap endpoint:
Client > graphql server > rest endpoint 1..3 > [sql server, document db, key-value store]
Where the graphql server takes a query, fetches and merges resources, and sends the result back to the client. (and as you mention, maybe rest and rest are external services)
If it's just:
Client > graphql > documentdb
(since your link mentions wiki...) - I'm not sure I see that graphql is a great benefit (but I'm perfectly open to that being wrong; as I said, I've only just begun looking at graphql).