From the server point of view, using GraphQL may not sound great since any single tiny update on the schema also implies an update on the single endpoint, rather than just one of many. Until there is an implementation of dynamic modularization, maintaining the integrality of a service is a constant pain.
Fairly enough, I think the tradeoff is balanced. GraphQL solves problems on the client side and yet it creates problems on the server side.
Meanwhile, I am more concerned about the future of GraphQL as http2 has arrived. With http2, the biggest selling point of reducing latency by using one single endpoint does not sound anymore. Then for the flexibility of limiting the scope of return, you can always specify it on the conventional RESTful API. So there is not much practical benefit. (Oh yes, there are some more goodies like subscription, but let's face it that not many use it) So to make GraphQL more useful, I think it needs a bigger evolution.
GraphQL's most significant latency gain isn't achieved through the use of a single endpoint to minimize parallel requests. As you mentioned, HTTP2 makes this a non-factor in terms of latency. The single endpoint part is only an implementation detail, and most follow it simply by convention.
Rather, GraphQL's major breakthrough in latency reduction is achieved through the use of the graph query language to completely bypass the multiple client-server roundtrip problem for interdependent data (i.e. having to wait for a user request to return the organization id, before being able to send a request to find some property off the organization the user belongs to). HTTP2 isn't able to help at all for this use case.