I am very doubtful there are going to be any optimizations that are going to make 16GB DDR3 1333 MHz RAM per CPU w/ 8 thread CPU obsolete by launch. I would not expect more than a 4x loss in efficiency per dollar spent compared to the upper 15% or so of miners, assuming AWS EC2 is not a clear way to go. My reasons for thinking this are:
1) The developers are biased towards CPUs, and I can't see that they are planning to optimized for GPUs.
2) There's no easily-available algorithm let alone code for optimizing GPUs for Equihash. If and when it becomes available, then you can get it. But I would not wait a month and miss out on a time when people are NOT using the GPUs efficiently. We don't have to run fast, we just have to run faster than each other.
3) If bandwidth is the limit, and GPU method is not made available, then the RAM+CPU cost per B.W. seems 3x more to get double results (DDR3 system verses DDR4).
4) If memory is the limit and the CPU can somehow make use of 32 GB with only 8 threads, despite the developer's goal of requiring 1 thread per GB, then 16 GB has loses only half.
Launch is supposedly only a month or two away, although a 3 month delay would seem normal because they have have not declared it is likely.