The existing attacks are about collisions: the attacker builds two certificates which hash to the same MD5, but with distinct contents. One of them is "benign" (it contains the attacker's name) and this is the one which the attacker sends to the CA for a signature. The CA signs it, because it is a perfectly valid certificate request (it comes from the attacker and really contains the attacker's name). Thanks to the MD5 collision, the signature is also verifiable when plugged in the other certificate. So the attacker got a CA signature on a certificate on which he chose the contents out of control of the CA. Details are a bit complex because the attacker must find a collision in such a way that the two resulting certificates are both "structurally valid".
What you ask for is something very different: you are asking about an attacker trying to generate a certificate with a hash matching that of an existing certificate, one that the attacker did not generate in the first place. This is called a second preimage attack and is generally harder than finding collisions. In particular, there is currently no known weakness in MD5 with regards to second preimages.
There are 2128 possible MD5 outputs (that's 1632, not 3216). On a general basis, if you have N input messages m1, m2,... mN, and you want to build a new message m distinct from all the mi but such that the MD5 of m is equal to the MD5 of one of the mi (any will do), then the best known attack method is simply trying random messages m until one matches. The average cost of that operation will be 2128/N.
There are probably much less than 232 SSL certificates out there, so this means that the attack cost will be at least 296, which is totally unfeasible with existing technology. So this attack will not work. Breaking a CA public key will be much easier (not feasible either, but still easier, by a factor about 1 million if the attacked CA key is a 1024-bit RSA key).
Actually, for certificates, the attack is more expensive than that, because a certificate contains the name of the issuing CA. Having a matching signature from the CA (e.g. through using the same MD5) is not enough for the resulting certificate to be of any use: it must also properly "chain" with the CA, which implies containing the CA name as "issuer name". So the N in the formula above is not the total number of certificates existing in the wild, but the total number of certificates emitted by a given CA which you have to target specifically. That's lower, yielding an even higher attack cost.