I don't get it. Any device maker has multiple options in those four configurations. All mobile chip makers seek to optimize "performance per Watt".
What parameters are Apple going to tweak to make it "exactly fit" rather than suffer the heartbreak of a "generic design compromise"?
Apple isn't custom designing the IPCores for the processor or graphics. I'm willing to bet that almost all of this chip is off the shelf stuff. What they can do is choose parameters like number of cores, cache sizes, CPU options (like you said, stuff that is available with existing suppliers) and also include additional units that they can license separately (e.g. earSmart)
The blog post mentions stuff like Frameworks and Apps. Unless I am very out of touch with SoC offerings these days, you don't get a MapKit.framework accelerator. You get things like H264 encoding/decoding and AES, or DSP functions which these frameworks can use. However, the faster performance you want with one function, the more locked in you get. An example of this is comparing encryption speeds on an Intel processor with AES-NI. AES is an order of magnitude faster than just about everything else. Everything else however, uses the same vector extension set that were designed specifically to speed up sub operations of common algorithms.
The other advantage they get by doing an in house chip is that they can choose the on board peripherals they want (possibly remove some chips that would normally be external), and the exact number of external busses they need for all the external peripherals, which can lead to a smaller, cheaper, more power efficient chip.