Sorry, calling that a block limit was an error by omission on my part. 2^32 yields a 50% chance of reuse. If we pick a sane security margin it's a lot smaller. Assuming I did the math correctly just now, 2^-32 only gives you ~2^17 blocks; dropping that to 2^-24 yields ~2^21 blocks.
Off the top of my head, NIST was suggesting something like 8GB as the working limit. It would depend on your risk tolerance and the application in practice I guess. For something like video you might not really care about exposing a few 8 byte blocks here and there where the exposure is one block XORed with the other.
An aside, personally I quite like TDES for the purpose of generating secure handles and the like. The larger block sizes of pretty much every other common algorithm yield URLs and integers that are more difficult to work with. 64 bits is a manageable enough length and you don't have to implement the algorithm yourself (at which point you'd have rolled your own crypto).
