According to an anonymous source of WCCFTech, AMD is preparing a 20nm-based graphics architecture that is expected to release in April or May. Originally, they predicted that the graphics devices, which they call R9 300 series, would be available in February or March. The reason for this “delay” is a massive demand for 20nm production.
The source also claims that NVIDIA will skip 20nm entirely and instead opt for 16nm when that becomes available (which is said to be mid or late 2016). The expectation is that NVIDIA will answer AMD's new graphics devices with a higher-end Maxwell device that is still at 28nm. Earlier rumors, based on a leaked SiSoftware entry, claim 3072 CUDA cores that are clocked between 1.1 GHz and 1.39 GHz. If true, this would give it between 6.75 and 8.54 TeraFLOPs of performance, the higher of which is right around the advertised performance of a GeForce Titan Z (only in a single compute device that does not require distribution of work like what SLI was created to automate).
Will this strategy work in NVIDIA's favor? I don't know. 28nm is a fairly stable process at this point, which will probably allow them to get chips that can be bigger and more aggressively clocked. On the other hand, they pretty much need to rely upon chips that are bigger and more aggressively clocked to be competitive with AMD's slightly more design architecture. Previous rumors also hint that AMD is looking at water-cooling for their reference card, which might place yet another handicap against NVIDIA, although cooling is not an area that NVIDIA struggles in.