On the client device, a synced Realm can be setup with an encryption key that's unique to the user and stored on the device keychain, so data is stored encrypted on the client. (related question: Can "data at rest" in the Realm Mobile Platform be encrypted?)
Realm Object Server and the clients can communicate via TLS, so data is encrypted in transit.
But the Realm Object Server does not appear to store data using encryption, since an admin user is able to access all the database contents via Realm Browser (https://realm.io/docs/realm-object-server/#data-browser).
Is it possible to setup Realm Mobile Platform so user data is encrypted end-to-end, such as no one but the user (not even server admins) have access to the decryption key?
Due to the way we handle conflict resolution, we currently are unable to provide end-to-end encryption, as you correctly deduced. Let's go a tiny bit into detail with regards to the conflict resolution.
In order to handle conflicts the way we do, we use something called operational transformation. This means that instead of sending the data over directly, the client tells the server the intent of the change, rather than the result. For example, when two users edit a text field, we would tell the server insert(data='new text', offset=0) because the first user prepended data at the beginning of the text field, and insert(data='some more stuff', offset=10) because the second user added data in the middle of the field. These two separate operations allow the server to uniquely resolve what happened, and have conflictless resolution of the two writes.
This also means that if we encrypt everything, the server would be unable to handle this conflict resolution.
This being said, that's for the current version. We do have a number of thoughts on how we could handle this in the future, while providing (some degree) of encryption. Mainly this would mean more work on the client, and maybe find a new algorithm that would allow us to tell the client the intent, and let the client figure out how to merge everything. This is a quadratic problem, though, so we're reticent to putting too much work on the client side, as it could really drain the battery.
That might be acceptable for some users, which is why we're looking into it. Basically, there will be a trade-off. As the old adage goes: fast, secure, convenient: pick two. We just have to figure out how to handle this properly.