I'm going to with hold judgement on the Kepler series, but due to pricing I'm already passing on the 7xxx series of cards and my 5870 is looking better and better each day.
I wish I could sit it out but my GPU won't allow it.
I think GK104 is their GTX-660, not 680 (assuming they don't change their naming convention from 40nm). It provides excellent performance per watt on a reasonably sized die, and will hold them over until GK110/680. Fermi had problems becsuse it was a new architecture on a new process. Its large size made it highly susceptible to imperfections on the wafer resulting in low yields. Those imperfections were a result of the newness of the 40nm process. Hence, large die is not a problem in of itself, the problem is a large die on a new/error prone process.
So I bet NV is releasing their "middle product" GK104/GTX660 with its much smaller die first, that will give them the time they need to 1.) let 28nm mature and 2.) design/optimize a big chip to decisively take the performance crown. In a way, that makes GK104 the "pipecleaner" product, does a good job carrying the banner, and giving time to produce the real king.