First off, a short comment on legality: Linus views binary drivers as legal ONLY IF they were "ported" to Linux. His view is, if they were written for the kernel from scratch, they're derivative work. Otherwise they don't derive from GPL code and (though undesired) are considered legal. Note that this "Linus" I am referring to may have no actual relation to the "Linus Torvalds" person who wrote Linux, it's just what I think I remember him saying.

Now that's out of the way, the first curveball that hit me after Win32 world was the idea of no stable ABI. First let's get some misconceptions out of the way:

1) Windows doesn't have a stable driver ABI.
Ask NVidia. Ask Intel. Every service pack that can potentially break a driver means that the ABI has changed! How many times did you have to update a driver after a service pack? How many times did a driver "require" a new service pack to install? Yeah, I thought so.
2) Distributions have "stablish" driver ABIs.
SLES and RHEL have got reference kernels with stable ABI calls that's well documented AND YOU HAVE THE SOURCE. It's much simpler to update drivers for Linux than for a new Windows service pack since the Enterprise versions normally stick to the ultra-stable stuff.

But, noooooo! I hear. Screw this stable stuff, stable stuff is old! And that's the crux of it. On a desktop, you don't necessarily want 99.9999% availability, you'd rather have 20 more features, but a non-serious hang once a month. The difference between the Windows model and the Linux model is: YOU NEVER GET WINDOWS KERNEL FEATURE UPDATES! You have to pay for them! And wait years and years... Yah, the service packs add features, but to most people they are considered bugfixes ;-)

So that's why the kernel shouldn't have a stable API (let alone ABI) for drivers. It prevents big architectural decisions from being taken at the opportune moment. With the kernel under git management, it's even easier to maintain out-of-tree drivers (case in point: alsa). What developers who clamour for a stable ABI don't understand is that there _is_ one! Any fork of Linux they make can be as stable as they want it to be. If they want to benefit from continual improvement, they _have_ to make some sacrifices.

Second-to-lastly, the guys biting of the short stick here are the distro maintainers. Again, there are many options: follow the source route like Gentoo, allow only free hardware like Debian, rely on the user base to spin driver packages like Fedora, Ubuntu and OpenSuSE, or go the stable route like SLES and RHEL. If they really want a stable video ABI so NVIDIA's drivers are even easier to install, get THEM to work together on it. If it's good, it'll get in to the kernel. If not, it won't.

Lastly, try the following experiment. Buy a shiny new Windows XP Home (or Pro, for that matter) CD, and install it on your brand-spanking-new hardware. Now, try to write a USB driver using only software that you don't need to pay for, and for which you'd own the binaries after you'd written it. Finding it difficult? If you got it right, PLEASE TELL ME FREAKING HOW, SINCE I DON'T KNOW! Next step, PCI driver... but hey, Rome wasn't built in a day!