They aren't Dell, so you aren't seeing Dell's much higher volume of Windows PC's. Most copies of Windows XP are sold with the hardware, not separately sold, as with the annual upgrades to Mac OS X. Amazon is clearly doing better on the Apple side of things, both hardware and software, than some vendors.
Another problem with these stats, not evident here, is that many copies of Linux are running on hardware sold with a Windows license. The three desktops that I am switching between right now are like that.
And yet another problem with such stats is that there are many brands of Linux sold or downloaded, and that each copy of a Linux distribution is licensed to be installed as many times as you like. So the original sales of a single Linux distribution are a much smaller percentage of the installed base than with Mac OS X or Windows.
And conversely, 80-to-90 percent of the computers in China and India are running pirated copies of Windows on computers that were originally sold with a disposable copy of Linux.
And yet another problem with such stats is that there are many brands of Linux sold or downloaded, and that each copy of a Linux distribution is licensed to be installed as many times as you like.
Multiple installations of a single copy of Linux is common for servers, but I believe the opposite scenario is common when Linux is installed on a single computer. When I'm installing a Linux server, I usually download several different distributions and test all of them. Then I pick one, install it and dispose of the others. I'm sure other Linux users do the same, and those unused downloads tend to inflate Linux statistics.
There is no universally agreed method for measuring OS usage share. But I think the best way to measure desktop OS usage is according to user-agent statistics on a broad range of general interest web server log files.
Linux is excellent for servers and embedded devices. It's the best OS for those purposes. But it's lousy for a desktop environment, especially for the typical home/business computer user.