I think the issue boils down to cost.. But just to help us with terminology, , I suggest a nomenclature of what I mean by the different terms.. (these are not terms I invented, but rather gleaned from among the first books on platform based design)
To start at the top (pun intended), System Design essentially is (and should be) hardware independent. If you study the famous "Y" chart, the three independent axes are Behavior, (also called as system specification or system design), Architecture (the hardware/software combination that executes any application) and Implementation that binds a specific behavior to a specific architecture. In this context, behavior is and decided independent of the implementation vehicle and then subsequently mapped onto various architecture choices to decide which architecture best suites the needs of that System Design.
This is in my mind, the ideal top-design design approach.
However, consider a scenario where a System Design company that goes through this process for their first design, and arrives at a decision on the "suitable architecture i.e. platform".. for that design.
If, for the second design, they are to do the same process and come up with a "different" platform, the cost structure to acquire, train, deliver and support products on a platform (outside of the core IP costs itself) are so high, that companies are loathe to re-invest that cost again for another design.
So, even if they don't throw away the whole architecture (i.e. processor(s), + OS(es) + IP + on-chip-communication structure), they tend to stay with one processor/OS combination and tweak from there.
Given that the processor costs (again, I talk about IP + training + delivery + support) are so high, the migration away from a specific processor is pretty hard to do, economically. Not that it can't be done, but it is usually not attempted.
In fact, if you study the system companies, they tend to have an in-house processor choice that is pretty consistent (ARM, MIPS, or Custom (tensilica, etc));
Even among the micro-controllers they tend to standarize on one (8051, Zilog, etc.)
Recognizing this, it makes sense for companies to "leverage a platform" of previous processor/OS/BUS/peripheral combos and that is what is a platform.
In fact, if you consider the nomenclature for Platform Based Design, they tend to be
"top-down" - generate the architecture that is best suited for the behavior you start from,
"bottom-up" - Find adjacent markets for the platform that you already have in some way, shape or form
"middle-out" - make some selections the same, but let others be decided based on context.
"Configurable" - where an on-board FPGA component will let you select the peripherals to go on the FPGA part and either the processors are hard (ARM as in the case of the Xilinx announced platforms) or soft processor.
Its just that the cost investment to a do a top down platform is so high, you have to leverage it across many design starts. Terminology wise you either have
"Platform" design or "Derivative" design, where the derivative uses a previously instantiated platform.