Seems Nvidia might have underestimated both AMD and overestimated Samsung's node.
A really bad combination potentially.
So let's look at what we know:
GA104 is the die that is currently being used for the 3070.
Normally Nvidia uses XX104 as their 70 and 80 class cards.
It looks like Nvidia may have been forced to move 3080 to a GA102 die, where the 102 is normally their biggest die, normally their 80ti class cards are reserved for x102 dies. This was likely done as an early reaction in case AMD had something good with RDNA2 based on projecting performance from RDNA1. This is also likely why the 3090 exists as a hail Mary halo card just in case 3080 is beaten so they can retain power crown. This would also make sense given the 3000 cards being pushed to their limits past efficiency to gain the extra performance.
Nvidia recently cancelled their 3070ti which would have been the full die version of GA104 as it is completely outclassed by Navi21 and as such wouldn't make any sense.
There have been rumours just recently that 3060 was being moved to GA104 instead of GA106, in other words being brought up a die to become a more powerful card. They would likely not do this unless they were afraid of being outclassed by Navi22 etc...
Then we hear rumours of a potential TSMC 7nm refresh of ampere, you don't do that unless you think your competition is too close for comfort. Of course they at the same time be unhappy with Samsung yields/performance, we will see what happens.
This brings us to larger memory variants being cancelled, they were definitely real and were listed on AIB roadmaps/docs. If they are cancelled it will likely be because Nvidia wanted to charge 150-200 more for them but now can't as AMD will be too competitive on price/performance so they likely cancelled them and are planning to refresh next year with supers maybe on TSMC 7nm