I agree that it is very sensible for AMD to focus on mobile and server. They rely on the likes of Global Foundries for their manufacturing capabilities, and while it is nice to see that they have been able to do some manufacturing at TSMC, it mostly stays the same. GF is well known for delivering late and not really have the performance they promised on paper years ago. A recent example is their scrapping of their fabled 14XM process, something they promised a good while back and never delivered. Actually looking a lot better with the licensing of Samsungs FinFET-process, that might actually work. I don't know why AMD sticks to GF like an abused spouse, maybe some VP:s are related or something. Make no sense from a business standpoint.
I don't think the Bulldozer architecture was a mistake really, it was just too late. First revision was obviously an early buggy alpha silicon, the second (Piledriver) was more OK and the third that's in Kaveri (SteamrollerB) is also an improvement. But also late to market.
As for the future, who knows? Excavator seems to be the end of the line for the Bulldozer arch, should hit sometime next year? Probably good for servers, not so much for desktop. The ARM stuff will probably never make it to desktop, the software isn't even close really and is still a problem for servers.