Nvidia has huge R&D costs this generation. We know of at least 6 chips - GP100/2/4/6/7/8. It takes a lot of money to design and tape out a GPU die, especially on this brand new Finfet process. They are integrating three different memory standards, two of which are brand new - G5X and HBM2. Granted, they have some pretty high margin products in Tesla P100 and 1080 FE, but they are going to need it.
AMD, on the other hand, have a very streamlined approach. Just two dies, and both are pretty small. P11 is similarly sized to SoCs, for which 14LPP is already pretty mature. Moreover, they will be re-using this tech for the console wins. PS4 Neo is 36 CU as well, it almost seems like a copy-paste of P10. I bet millions of P11s have been snapped up by the likes of Apple and Dell to go into their laptops. P11 simply has no competition and won't till at least the end of the year.
All of this means that desktop gaming GPUs are going to be much more affordable. AMD can afford to make less profit per unit because there's far less R&D costs to recover. And going for a high volume market, even that is spread over a large quantity - so each GPU has a much lower R&D cost attached.