You generally get what you pay for. With Dynamixels, you are making a trade-off of getting a fully functional and tested actuator out of the box, versus spending a lot of time rigging up your own actuators, and probably dealing with numerous engineering efforts. If you are making a million robot arms, it probably makes sense to spend significant engineering effort to design custom actuators. If you are making only one, why not leverage the engineering effort that Robotis has already put into Dynamixels.
If you open up a Dynamixel and take a look at what is inside, you'll quickly realize the prices are not crazy. There's a high end motor, numerous custom gears, a custom case, custom driver board -- furthermore, someone built this product, and then a retailer handled import/storage/etc so that you could go out and have it delivered via Next Day UPS if you so choose. And if something goes wrong with that servo, you have a company to call up and get a replacement from.
The best hope of cost reducing these components is huge volume. Figure out a killer app, and order parts for a million 7-dof arms, and you will almost certainly get a nice quantity discount. Of course, you probably wanted to make a profit on selling that 7-dof arm with your software, right? Or at least, keep yourself employed? Well, then your arm is still probably going to retail for more than $2k.
Also, I'm not sure how we commercialized cars, but they cost more than $2k. Maybe the robot just needs to generate more value...
