Even in good times the GPU business has always been about breakeven. That's still better than the CPU business is doing though. Even if Carrizo turns out to be a decent product they still have to deal with the reality that consumer PC sales are still slumping plus Bay Trail now dominates what's left of it since Intel is selling it for basically nothing.
AMD really needs the Zen architecture to be a big hit. Sadly it's still a ways off, which is really leaving their performance lineup very long in the tooth, considering it really didn't put up much competition when it was new.
AMD has been in need of a completely new architecture that bucked the current trends they were working with. Intel had it's field day when it had the Pentium M, basically a shrink of the Pentium 3, which heavily influenced the Core architecture. They went on a tear with all of that, and Nehalem was just as much of a major step from the Core era, and the retooling of Nehalem into Sandy Bridge was phenomenal. Since then, Intel hasn't really really gone back to square one to really make the next big step, because they don't have to. They are reiterating and improving, but they haven't provided a significant leap since Sandy Bridge.
It has been said that Zen is such an example of a drastic revisit in their microarchitecture design, and is also a return to greater threading performance as opposed to leaning on weaker but more numerous cores.
If they can take a page out of Intel's playbook, look back to previous champions like K8 and reincorporate what worked so well there but of course improved with today's knowledge and technology, they could really have a hit on their hands.
I remember it was always hard for me to "switch camps" at different times: first to AMD during the K8 craze and enjoying overclocking and ultimately buying into Athlon; and then finding it a little hard to accept that Core2 was far and away better than anything AMD was offering but ultimately getting into a C2D to overclock. I've stuck to "the Intel camp" since then, never finding the AMD performance to ever be worthwhile compared to what I could get and afford with Intel. Granted, I have only made one jump since that C2D, to a 2600K, and am looking to hold onto this build until Skylake. And really, Skylake is just to get the latest technology I have been missing out on, I find the Z68 long in the tooth, not my CPU (especially OC'd). Performance charts still prove that to be true. But damn, it is now becoming time for PCIe 3.0, and the likes of native m.2 PCIe SSD support and SATA-Express would be quite handy.
My long-winded point, Intel is now simply cashing in on simply being able to get people to upgrade based on new technologies and ports (USB 3.1 Type C is freaking outstanding ), barely doing more than iterating on past CPU architectures.
If AMD can not only do a complete architecture overhaul from the ground up, but also succeed in making it a market contender, we all win, we all win very handsomely.
On the GPU front, I don't expect the R9 300 series to ultimately shake things up too much, but if it does compete fairly squarely with Nvidia's Maxwell2 in both performance and power efficiency, that's one hell of a starting point. Even if that doesn't change their market share, if they are basically neck and neck with Nvidia again, that at least demonstrates that they can continue to push their architecture. Hopefully it does bring them some additional revenue, because more investment in the GPU division could provide enough of an R&D budget to not only push GCN, but ultimately incorporate what worked into a new architecture, because it is time for that. I think they're running up against the limits of what one architecture can truly accomplish with only minimal improvements (in the grand scheme of things).
A die shrink to 14nm obviously gives them more room to give one more mighty push, and I expect that AMD's next generation will compete squarely with Pascal. I'm fairly confident that Pascal will ultimately shape up to be more of a Tick than a Tock, mostly serving to be a die-shrink that is perhaps re-tooled and improved upon slightly but not a drastically different architecture design. The Maxwell architecture was ultimately designed for a 20nm node or smaller, but the node improvements for GPUs fell back with delays. If this new version of GCN matches or bests Maxwell, then the die-shrink from both companies should keep things mostly around that same toe-to-toe matchup. After that, however, I really hope AMD is pushing to move beyond GCN or so radically reinvent it that may as well be new.