There is a chance in the next 10 years the discrete GPU market will just be gone. Every one will need some amount of graphics capability but once the integrated parts get good enough as more and more transistors and capabilities find themselves onto the CPU we may well find that discrete cards don't offer much advantage.
Intel could build a discrete card, they have certainly tried all be it via a somewhat odd direction. But Intel doesn't just lean on its fabrication technology, it also leans on its IP with x86 and combining the two does take them down a route that none of their competitors can really go. Graphics is a special mass market case of highly parallel computation depending on a lot of floating point and various fused instructions and some quite specific computations. Its certainly possible to write a graphics pipeline as a CPU program, given enough compute resources it should be possible to create a generic version of a GPU that has a lot more applications. If you have ever tried openCL/Cuda programming I can tell you its not pretty, its very specific and you have to change your software a lot to make it work.
I think Intel is on the right long term path, integrated for the usual desktop world and the future of mass parallel with their current Phi research projects, both of which have the potential to completely change the discrete GPU market completely. They don't just want to compete in the market, they are attempting to put themselves years ahead of any potential competition.