How do you figure that? You think they sold 4 times as many 8000+ GPUs than all the previous combined? Even accounting for retirement of older hardware.... many (most) linux users aren't in it for the top-end graphics anyways, which means that most are happy with a slightly older card. My guess would be that on linux, the 8000+ cards are VASTLY outnumbered by older cards.
Heh... And how can you get that with proprietary drivers? You're NOT supposed to be shipping those blobs with the distribution- period. Doing so breaches your GPL license grant on the Linux kernel. You can, post install, do it as an end-user (since there's nothing keeping you from doing whatever you want with things...)- but a distribution vendor can't do that as they're publishing/distributing, which is where the GPL takes on full force. You don't want to go there.
Since you can't distribute a working 3D distribution for an NVidia user- they have to download/install, post installation, the working 3D support, unless you use Nouveau drivers, NVidia can't give that experience. Only Intel and AMD (after a fashion, depending on which model of card...) can give you that. Because they're FOSS.
NVidia loses because of that piece. If the FOSS support were more there with AMD or Intel had actual credible 3D chipsets, there'd be a differing story with NVidia in this space.
Hm... Considering that they've just added hardware tessellation for GL3.2 on fglrx, I'd say that while they're not as stable as NVidia's blob, AMD's is on parity or slightly ahead right now. Main reason I don't use AMD's stuff is more because the driver's aren't as stable as the Linux ones right now.
Mesa's the only one playing catch-up in the sense you're talking to there. Moreover...here's a hint: Unless Ryan Gordon, LGP, myself (as part of LGP or as an indie port...), or someone like Uningine's team does something with OpenGL 3.2, you're not going to see a lot of that being used for at least 6-12 months yet to come. 3.2 support's sort of nice, but it doesn't impact the card's performance, driver usability, etc.
You may forget that it is even possible to install the nvidia binary in live mode (this partly is working for fglrx too but with lower success rate).
I don't know how the NVIDIA usage breaks down, but I do know that you shouldn't be relying on those numbers. They represent the windows gaming market, which is obviously going to move to newer/faster cards much earlier than the larger general purpose market. I have no idea how Linux OS affects the numbers - there might be more people throwing linux on old machines, or there may be more users who are tech enthusiasts and upgrade their machines more often than the typical windows user.