If Mailpile is the MTA, and the NSA is following the connections to Mailpile's servers they can use time correlation to find out the metadata.
Metadata is hard to hide^.
Anyway, as the other posted mentioned, the correct solution is definitely full security across the board, with full encryption MTA->MTA. That needs to happen so it can't be sniffed on the wire. But of course full security between Alice's MTA and Bob's MTA is ultimately pointless when Bob's MTA (e.g. Gmail) is sleeping with Evan (the NSA).
The actual message contents need to be decrypted with keys that only the recipient has access to, and GPG is as good a solution as any for that. You can only trust what only you have. Don't trust Google. Don't trust Mailpile, either.
^ anon.penet.fi was an OK solution, but still was subject to possible timing attacks (and legal attacks since a table exists SOMEWHERE in the universe that correlates your name to your nym).
A nym.alias.net solution is better when you're chaining remailers and using random timing delays.
Deadrops are better, post your message anonymously to a newsgroup/forum/etc via a Tor or other anonymous connection, and your recipient does the same to retrieve the message anonymous. There's no metadata there to capture.
The server decrypts EM2 revealing EM1 and some plaintext metatdata arbitrarily specified by Alice. The MTA random-delays to keep them all out of order, and sends it on to Bob with encrypted body, whatever Alice felt like putting in the plaintext metadata supposedly representing the prior travels, and no attempt to disguise the MTA's IP, etc..
NSA now can see Alice's post to the MTA, but none of the mails coming out of the MTA match the text of Alice's email. The best the wiretapper can do is decide "someone in set A sent to someone in set B", and maybe apply statistical analysis. It is easy for the wiretapper if there are only a few people using this and hard if there are a lot.
The objection will be, but NSA/FBI etc. will trojan the code, coerce the secret keys, make the server deceive users. So the server owner would like to put in something that makes it all unavoidably, and conspicuously break if the code is compromised. Securing the latter behavior remains a problem when the adversary has, one must assume, physical control of the server.