Remember that even with the potential for supposed GPU-acceleration, (which as of yet still remains a pipe dream given the difficulty of programming them even with things like CUDA, though Fermi might change this) you're still looking at testing a LOT of passwords for a brute-force attack. If someone does have both upper/lowercase alphanumeric, and still has no special characters, that means the total combinations will be equal to 62^n, where n is the number of characters; even a "weak" 6-character password would mean 56.8 billion combinations... And remember that it takes far more than one single-precision floating-point operation to test a password.
As for hacking attempts releasing passwords, I still find it deeply disturbing how many (even large) websites still actually store their passwords in an unencrypted form. Most decent sites and services know that such a method will never, EVER be secure, and instead only use hashes; i.e, if you broke into the "password" database, you'd only get bunches of fixed-length binary strings with completely random distribution, from which you would NEVER be able to ascertain the original password, because the hashing process, in effect, is "one-way;" when a user submits their password to log in, the server applies the hash, and sees if the HASHES match. This scheme leaves brute-forcing as the one and only method of cracking them.