When people talk about the benefit of HBM, they talk about ditching traditional memory for HBM. When you combine it you lose most of them.
Hmm, not always. There's been quite a lot of speculation, and rumour, about AMD putting a (much smaller) amount of HBM on a APU to work (via HBCC) with general DDR4 to alleviate the common bandwidth problems APUs have working against DDR alone. Putting memory on or near the die as a compliment to larger system memory is not an unheard of idea more generally.
But I agree GDDR6 seems more likely, and the option floated with HBM2 seems to require some subtle calculus based on the public info we have. I would assume if someone did choose this that they did their homework though, in terms of costs and cost scaling.