The Texas TACC Stampede will be configured with 6,400 Dell DCS Zeus compute nodes, each with two 2.7 GHz E5-2680 Intel Xeon (Sandy Bridge EP) processors. It is targeted to have 6,400 PHI. They have connected 1 PHI for each Sandy Bridge EP. The PHI bandwidth probably makes it impractical to connect more than 1 per CPU. I don't know what the real phyical limit is.
The population of sockets is not, however, limited to servers. There are workstations too. Gamers? Is Intel working with game developers? There will also be substantial price elasticity. The PCI add in board will find additional installations as they are able to drop the price. Workstations are also targets.
The Intel C and Fortran compilers will make it easy for people to build code that take full advantage of the PHI and capture near 100% of its compute capability. MATLAB, simulations, modeling, ...
I have an Intel Core i7 Ivy Bridge desktop and I have been watching for PHI to hit retail and drop to my price point. I am cheap so it will be awhile but I can put it on my Wish list.