He didn't said that (big) Vega won't ship with HBM2, but that AMD has to either make (big) Vega faster or make it cheaper.
So what exactly does that have to do with anything?
Your friend doesn't have much of an idea what he is talking about, does he? So whats the cost of HBM vs HBM2? whats the cost of the passive interposer? But more importantly whats the cost of 12 GB of GDDR5X vs 8 GB of HBM2!
I dont think people like you realize how cheap it is to manufacture CPU's/GPU's ( you can bin/salvage them so well) and how expensive a massive amount of a extremely boutique memory type is, GDDRX5 is used in far fewer products then HBM2.
But consider this Vega is smaller then Fiji, with a smaller interposer and fewer memory stacks using a technology that has a vastly larger amount being manufactured. So even if Vega performed as poorly per as fiji it is still not going to be a worse product financially.
This is because Vega already supports GDDR5X and switching from HBM2 to GDDR5X would cause the least delay.
You really have no idea, so what your saying is that AMD has wasted something like 15-20% of its die space to package a 384/512bit GDDR memory interface because you certianly cant use the HBM interface! Or your other option is you go all the way back to RTL and change the memory subsystem. So whats the cost of that, consider opportunity, development and the impact that has on future development?
Then we move onto they delay, what delay exactly? You realize your stupid HBM to GDDRX swap is not a 3 month delay, more like 9 to 12.
But what your financial expert friend has really missed is the part where AMD has said they have used there priority access to prepurchased millions of dollars of HBM2 up front.
But what i find even funnier is, that GCN is still more advanced then pascal in terms of compute, Vega is bringing big updates in the area's that NV has clear advantage ( ROP's, ROP cache, frontend, increased clock). But then on top of that they are bring the follow on ALU architecture to GCN.
Go and read any of the GPU ALU paitents from the last 2-3 years from AMD, if any of it appears is NCU then GCN's advantage extends.
You friends logic doesn't stand up to even the simplest of scrutiny.