Simon-- There is a lot more to cryptography hardware than just making stuff go fast. Any kind of co-processor array could handle the job -- doesn't have to be DSPs. For unsigned integer math, the 486 DX4/100 is a perfectly decent chip, and prices are plumetting. If all you want is to make RSA go fast, you can rack up as many of these as you want, and do your work at Ethernet speeds (which should be more than fast enough for the forseeable future.) Also, you get to use your old familiar (and inexpensive) programming tools to write your code. It also turns out that when you're actually using cryptography to do valuable stuff, a lot of other concerns are just as important as speed -- in particular, physical security over the private keys. Most server-oriented cryptography add-ins are highly tamper-resistant units that can only be re-keyed when some sort of dongle is attached. Said dongle can be kept much more securely than a whole computer. Although other threats still remain, this substantially reduces the threat of key compromise. Incidentally, good hardware of this sort uses ASICs that give substantially more bang for the buck than the DSP solutions you're suggesting. Certainly messing around with DSPs is fun -- I've done it and enjoyed it -- but I don't think they will ever be significant as cryptographic co-processors.
That's another question. A DSP chip can also be used for crypto - yet sound cards and nexts aren't ITARed, and aren't really considered dual-use. A Modular exponentiator isn't a crypto device (hey - it's a bignum accelerator for Mathematica). Now, if I had a pipelined WSI chip capable of delivering one result per cycle, I could think of some useful applications, but ...
Simon