My theory (admittedly without evidence) is market segmentation, on both ATI's and NVidia's parts. It's something that has been done for years in the tech community, across many different kinds of products.

In effect, given the costs of production, it would be a lot cheaper for both ATI and NVidia to make a single GPU, and use binary drivers to enable/disable additional pipelines, texture processing units, etc, than it would be to actually make a series of different GPUs that have those capabilities. It wouldn't surprise me much at all that other than actual clock frequency & RAM speed, the only difference between the $100 cards and the $400 cards (assuming the same family of GPU) is an ID somewhere deep on the card that allows the driver to determine how many pipelines and additional features to enable. Consider this the difference between the 'student' and 'pro' versions. :)