Yes, WereKitten's explanation was good.
It remains to be seen whether the overhead of x86 is a good enough tradeoff for the extra programmability. Current GPUs are quite inflexible and difficult to program for, and have a lot of dedicated hardware that can't be repurposed, but that makes them efficient on power and die size terms. GPGPU is such a niche market at the moment that Larrabee's first iteration would not be competitive in the market right now.
What Intel would eventaully do with it is integrate lots of Larrabee cores onto the die of a CPU. So you would have one chip that would be a CPU and GPU. AMD's Fusion (Llano) is the same idea but using their processors and Radeon GPU 'cores'.