Neither.
The ever-increasing costs of mask sets for any kind of custom chip and
the ever-increasing cost of getting decent cost and performance out of
FPGAs will lead to more hybrid pattern-able custom chips: SoCs with
hard-coded processor / DSP / memory / standard peripheral cores but
with final metal mask(s)- or fuse- or flash- programmable logic cell
and/or analog array areas for specific applications.
You're already seeing that, though DSPs not so much. Everyone has
soft core and hard core processors of various stripes depending on
needs. RAM/ROM in single and dual ports have been everywhere for a
decade. User flash is available on several models. The FPGA fabric
just begs to do the DSP type work so I don't see too much there.
Peripherals, except for ubiquitous things like USB, won't find their
way into hard macros either. Hardware accelerators, such as DDR, QDR,
and other SerDes interfaces already have.
Vendors that
cover as much application / volume space as possible with the least
investment and best service (models, tool chains, prototype
turnaround, application support, etc.) will be the big winners (as in
every other generational transition of semi-custom silicon product).
Think Microchip (or higher-order) with FPGA and/or programmable analog
blocks.
They might be the "winners" but there will be many. The real money is
in the niches. 'X', 'A', and 'a' have a *pile* of money tied up in
the things you cite.