It's not just about power, bandwidth as mentioned is also a feature plus being able to make the pcb's shorter. Gddr adds a lot of complexity to graphics cards pcb layout compared to having it all bundled onto a small area around the gpu die.
As for power, there's this video that talks about it, its not a myth at all:
The bandwdith is a bit of a myth though. The 1080ti has more bandwidth than Vega64, and GDRR6 will increase this a lot.
The PCB complexity is a trade off, and I wouldn't even say complexity but size. GDDR needs more PCB space, and there are more visible connections. But mounting GDDR moduels is very simple, failure rates are low and if a module is not properly seated it is easy to repair after testing. If the problem lies within the die then that die can be used for a lower end SKU.
With HBM, the PCB is smaller and it looks simpler but the complexity is moved to the HBM chips, the interposer and the GPU die. In total it is much more complex, much more fragile, far harder to manufacture and when there is a failure the whole interposer, HBM and and GPU become a paperweight (imagine a Vega64 with only 1 working HBM stack, lol).
This is kid of of obvious otherwise HBM would have been used a long time ago and would be far cheaper. The whole concept of stacked chips and interposer is very obvious, and have been used since the late 70s (TSC technology wa patened in 1962!). But the process complexity has prohibited the technology until recently.