The thing with the VRAM situation is NVIDIA's in a bind, and I wouldn't be surprised if AMD is going to be in a similar situation. As far as I know, neither Micron nor Samsung have a 32Gb (4GB) GDDR6 chip. And while I can find plenty of articles in 2018 saying that GDDR6 is capable of 32Gb and there were supposedly plans to make it, so far neither company has said anything about it (or my Google Fu isn't that strong at the time of this post)
Memory in video cards is also basically n-channel, which n being the number of VRAM chips on the card. And since each channel is 32-bit, a 192-bit memory channel means 6-channels. On top of that, for maximum performance, each memory channel should have its own memory controller. You could do something like segmented memory to add more VRAM, but nobody wants that anymore since the GTX 970.
This is also on top needing to route all those traces on the PCB, to make sure those traces can reliably carry the signal.
So these are your problems:
- Maximum capacity per chip still appears to be 16Gb (2GB), which obviously means even more chips for more memory
- To add more memory, you need to keep adding more memory controllers. This eats into die space, adds to power consumption, among other things
- And to support more memory controllers/channels, you need a more complex PCB for routing and reliability.
While AMD probably can get away from the second problem with an off-die memory controller, it still doesn't resolve the first and third problems.