I can't update my original post now, but I stand corrected. I "remembered" uploading the private key with the CSR last time I went through the process, but must have misremembered. Well, that's some good news I guess. The NSA could still easily request the private key from the company, of course.
You don't share your private key with the CA.
A certificate is an attestation (signature) by a CA's private key that a given PUBLIC key is yours, that anyone with the CA's public key can verify.
The CSR does not contain your private key.
NONE OF THIS MATTERS. This is not about bulk-decrypting SSL, though I'm sure NSA does that when and where they can, too. This is about coordinated, automated, integrated methods of transmitting the plaintext.
Why should they bother getting a key and scraping gmail's payloads when they could just have Google give them an API? Furthermore, this method would continue working perfectly even ifwhen services switched to ephemeral key modes that provide PFS.
Getting a root's key does not enable them to decrypt the traffic. Getting the server's SSL private key USUALLY means you can retroactively decrypt, but it is possible (though uncommon) for servers to be configured to use ephemeral keys (EDH modes) which provide perfect forward secrecy (that is, the property that the session key can't be recovered even with later compromise of the server's long-term key).
Interesting sidenote: AFAIK there are no widely-supported (TLS 1.1 or below) methods of mitigating the BEAST attack while enabling PFS - those modes are TLS 1.2+ which isn't widely spoken yet.
This is all irrelevant though. The issue is with the tech companies giving them the plaintext data themselves. No amount of transport-layer encryption helps with that.
You have to stop using US services.