You've got to stop thinking of it as a GPU and think of it more like a co-processor.
First of all, AMD isn't going to force you to buy a built-in GPU on all of their processors. Obviously the enthusiast market is going to want huge 300W discrete graphics rather than the 10-15W integrated ones. There will continue to be discrete CPUs, just like there will always continue to be discrete GPUs.
But this is a brilliant move on AMD's part. They start with a chunk of the market that is already willing to accept this: system builders, motherboard makers and OEMs will be thrilled to be able to build even smaller, simpler, more power efficient systems for the low end. This technology will make laptops and netbooks more powerful and have better battery life by using less energy for the graphics component.
Now look further ahead, when AMD begins removing some of the barriers that currently make programming the GPU for general-purpose operations (GPGPU) such a pain. For example, right now you have to go through a driver in the OS and copy input data over the PCI bus into the frame buffer, do the processing on the GPU, then copy the results back over the PCI bus into RAM. For a lot of things, this is simply too much overhead for the GPU to be much help.
But AMD can change that by establishing a standard for incorporating a GPU into the CPU. Eventually, imagine an AMD CPU that has the GPU integrated so tightly with the CPU that the CPU and GPU share a cache-coherent view of the main system memory, and even share a massive L3 cache. What if the GPU can use the same x86 virtual addresses that the CPU does? Then...all we have to have is a compiler option that enables the use of the GPU, and even tiny operations can be accelerated by the built-in GPU.
In this future world, there's still a place for discrete graphics -- that's not going away for your gaming rig. But imagine the potential of having a TFLOP-scale coprocessor as a fundamental part of future sub-50W CPU. Your laptop would be able to do things like real-time video stabilization, transcoding, physics modeling, and image processing, all without breaking the bank (or the power budget).
But before we can get to this place, AMD has to start somewhere. The first step is proving that a GPU can coexist with a CPU on the same silicon, and that such an arrangement can be built and sold at a profit. The rest is just evolution.