I really can't see Nvidia in the CPU market. There are other markets available to them with better margins. For one, it would be cheaper for Nvidia to buy VIA just for the license on x86 than to go straight to Intel, since the 2 would squabble away for years, i mean, they couldn't agree on a chipset license, so what are the chances of getting x86. The CPU market is very harsh atm, AMD is barely staying afloat, but its server market is very good (12 core Opterons near perfectly scaling in 4 socket systems). If Nvidia can't get an x86 license, they'll either have to create a custom platform or use something like ARM, SPARC or PowerPC, which still means they're tied to other companies. Creating a new core logic and getting everyone to follow it would also be an uphill struggle (motherboards, chipsets, NICS, I/O, etc). The only real choice is accelerator cards, so the push for GPGPU is a justifiable direction for them to take.
Technically it's all full circle, CPU's were integer units, then floating point units were introduced as accelerator cards for specific needs, then FPU's were integrated (partially) into CPU's, then graphics cards came along (which are massively parallel FPU's), and now both Intel and AMD are integrating the graphics units into the CPU. So Nvidia would be making very big FPU's with API extensions. Their concentration on proprietary software technologies is their downfall though, since developers would want to be able to program for any FPU rather than Nvidia only (and a specific model at that), it would be like writing software that'll only work on Intel processors.
There is the lucrative mobile market as well, which is where their Tegra and ION platforms come in. They can mix it up with Console graphics as well. There really isn't any point in them getting into CPU's when they can provide dedicated accelerator chips for a broad range of apps without the worry of dealing with cross license agreements and other legal issues over licenses. It would be nice if they came up with a new core logic for CPU's, but x86 is so entrenched in high end computing, they could never replace it. Thus we will be stuck with a 30+ year old architecture with strict backwards compatibility for that one high paying customer that can't be bothered to change hardware or re-write their software, for at least another 20 years, or until Intel goes under or opens up x86.