Windows spyware phoning home. Botnet trojan participating in a DDOS. More innocent is simply the design of the network stack. OS X uses the time-proven BSD stack, and Linux's stack is also good, but Windows XP's simply sucks, an attempt to fit a limited version of the BSD stack into the quite different Windows API (which is why Vista has a completely new networking stack).
Yes, any application consuming significant network bandwidth will leave less available for a speed test, but then that test isn't a valid measurement of bandwidth. Such applications should be terminated prior to running a valid test.
More innocent is simply the design of the network stack.
See my post #19. Such factors may come into play at gigabit Ethernet speeds, and I would be inclined to believe any statistically valid tests that were performed to prove such, but not at common <10MB broadband Internet speeds.