gritz .. Yeah, simulating them is very slow. It gets exponentially slower the bigger the net too. But these chips can be linked together. You can clock one on it's own @27MHz, but when you link more than one, it's a one off cost of half the speed. And as far as I understand, you can just keep daisy chaining more and more together.
bobc .. Have a look at some of the demo videos. I can't see these chips having any trouble reading current capchas .. Looking into the Human Brain Project now, very interesting.
On a side note, here
is a little neural net program I wrote a while ago (java, runs in browser)