From this page, I see a description of MS-Cache v2 as the MD4 hash of the concatenation of the "NTLM hash" of the password, and the username (converted to lowercase, then encoded in "Unicode", which means "little-endian UTF-16" in Microsoft terminology). The "NTLM hash" is MD4 computed over the password (there again in little-endian UTF-16). So the hash involves two nested MD4 invocations and is "salted" by the user name.
Rainbow tables are "just" precomputed tables (with a trick which saves a lot of space); they can crack hashes which were encountered during the table building phase, but none other. Therefore, you can build a rainbow table for MS-Cache v2, but only for a specific user name; the table would not be applicable to hashes for other users. Therefore, such a table would make sense for a user name which is often found in many domains, namely "Administrator", but this is still of limited applicability. Cain & Abel pretends to include (since version 4.9.43) support for generating and using rainbow tables on MS-Cache v2 hashes, but, of course, for a single user name at a time.
If you only have one hash to crack, then building the table would take more time than brute force cracking (building a rainbow table which covers N possible passwords takes time about 1.7*N whereas brute force cracking has an average cost of about N/2, so the table is worth building only if trying to crack at least 4 hashed passwords, all with the same username in the case of MS-Cache v2).
I did not find freely available, already computed rainbow tables for MS-Cache v2 and username "Administrator".
MD4 is an ancestor to MD5 and maps very well to GPU (even better than MD5) so your four GPU should be able to try potential passwords by the billion(s) per second. The "jumbo" version of John the Ripper is documented to support MS-Cache v2 with GPU (under the name "MSCash2"), albeit with some inefficiencies on the password generation side (to hash one billion passwords per second, you have to generate one billion potential passwords per second, and that could starve the most well-intentioned CPU). At the very least, this will give you some source code to play with, and a "reference implementation" which would be handy for tests.
A quad-core Core2 2.4 GHz CPU, without any GPU at all, should be able to compute 80 millions of MD4 hashes per second, with SSE2-enhanced code (estimation from an implementation of mine which did 48 millions of SHA-1 hashes per second), translating to 40 millions of passwords per second. This is already quite good, by password cracking standards, although of course the GPU will be vastly more efficient. It would be a nice programming exercise.