Judging by some of their comments in their FAQ and stuff they don't even understand how A* works, so I wouldn't put much faith in this thing. An AI coprocessor could be very beneficial in theory. Branch heavy graph searching algorithms could benefit greatly from specialized hardware. In Game Programming Gems(6 i think), there is an implementation of A* on the GPU which turned out significantly faster. A more generalized AI processor would be pretty cool, but I think it's going to be much harder to get market with one. Unlike Physx, which can happily do eye candy only so lesser machines can still experience the main parts of the game, I figure you wouldn't be able to rely on this AI thing and still give those without it the full experience. The industry is needing more talented AI programmers with more time to do AI than they do an AI processor, especially with duel and soon quad core at our disposal. If I had to choose I'd probably prefer a physics accelerator that let me do shitloads of hardware accelerated collision testing, which is one of the main expenses of much AI work. Traceline, TraceHull, etc. I think PhysX screwed themselves with their rediculous $300 price tag. I could see paying up to $100 at most. Likewise, if an AI chip hits market, it will probably be overpriced as well, ensuring a crappy market penetration, which in turn means less likely for developers to bother, which then even further gives gamers less incentive to get them.
p.s. The network card you speak of is this joke
http://www.killernic.com/KillerNic/