AMD also just packed flexibility and backup plans.
Take the HBM issues. Nvidia knew stacked memory could be useful and initially worked with micron, then realized HBM would better, but it wasn't ready for mainstream due to high costs, they smoothly switched to GDDR5x, and now 6. AMD however put all their eggs in one basket. HBM made Fuji short on Vram, delayed production, increased production costs and complexity, reduced supply. Nvidia by only using HBM on server parts meant margins were huge and volume low, perfect to get a new tech going.
I fear AMD are still looking at HBM but the future HBM just don't look that enticing compared to to GDDR6. HBM2 hasn't really scaled much in density or speed. The next HBM stabdrst on the horizon is supposed to be a slower lower cost lower power version for embedded, nothing for a high end GPU.
The whole GF thing is also kind of their fault
HBM hasn't really shone in GPU use - GDDR has really benefited from shrinking nodes and continues to at a rate HBM doesn't seem to mitigating the power and heat considerations and allowing evolutions of layout, etc. to offset the potential latency downsides. Then there is that HBM seems more suited in methods of access and dealing with data to compute type tasks and only certain areas of rendering/gaming use where largely GDDR seems more suited.
TBH I think its time to start looking at types of memory beyond GDDR but meh GDDR6 will do for now. But I called it a long time ago on HBM and was shot down because it was newer and so it had to be better right? :s