Buying ATI was a good move, spinning off their Fabs was catastrophically bad. Real men have fabs, that's a quote from Jerry Sanders, who kept the company alive, and created the successful AMD.
They desperately need to produce something after the K7, which is still what they have with the Phenom. It's a bad design, and pretty much always was, and only was successful when competing against an old design like the P6 (and even then, not always well) or a really bad design like the Prescott. Living off the success of your competitor's failures is a mistake, especially when it's Intel. Sooner or later, they get it right, and they have.
AMD is in this situation because of the failure of Hector Ruiz, and his inability to significantly change the K7. Tweaks are not enough, and they have to create something almost from scratch. Things like a powerful x87 unit are almost completely useless, and consume power and size, and add cost for no benefit. Supporting 3D Now! is useless, and wastes transistors (not many, but some). These are just examples of waste, and obviously there are serious design issues with the K7 that they have not addressed. The fact that the Phenom is roughly the same size as the i7, yet so much slower is a first for AMD. It's the first time they had such a significantly slower chip the same size as the Intel chip, and more than that, is significantly slower than the previous generation Intel chip. The current design, 10 years later, being so close to the K7 was catastrophic for AMD, and thank goodness for that. Anything else would have pointed to a lack of progress in the industry.
The AMD/ATI synergy is a good thing, and naturally it has some bumps. But, how else could you possibly recommend an AMD solution were it not for the 790GX? The processors aren't very competitive except when they don't matter, but the IGP is significantly better than the Intel IGP, thus giving AMD a significant advantage to offset Intel's in processors, in certain situations.
I doubt AMD has the resources, since they can't seem to develop a decent x86 processor, but I'd love to see them develop a new RISC chip, without the x86 instruction set impediment. I'd like to see them focus almost entirely on the server space with it, since getting a new instruction set for the desktop would be nearly impossible. A partnership with Dell would make it perfect (IBM wouldn't work because it would compete with POWER, HP wouldn't work since they designed Itanium, and Sun is being shopped around so has little relevance). Since decoupled processors are slower (more stages, greater misprediction penalty), use more power (all the extra transistors needed to decode variable length x86 instructions to RISC type instructions), generate more heat, and are larger and thus more expensive (except economy of scale helps x86 a lot), a new design could be appealing in the server space. The big "plus" for x86 is greater code density, and thus greater cache effectiveness, but this is untrue for the K7 since it pads the L1 cache to deal with the variable x86 instructions. If they can do it, they should. But, I doubt they will, they have been very conservative since Jerry Sanders left, much to their detriment.