With AD102 the only part available thus far is 4090 - and you'd assume that if Nv would be able margin wise to do this they'd re-use AD102 for 4080 at least, especially since that one is $1200 now. The fact that they preferred to make a separate AD103 die for this card highlights how expensive 4N process really is.
The whole "Nvidia is greedy" talk is baseless. Or do anyone here really think that Nvidia of all parties have managed to miss the end of ETH mining?
Samsung's 8N process is really at its core an 'enhanced' 10nm-class node, and TSMC's 4N is really a 5nm-class one.
In addition there were reports that NVidia had negotiated a deal with Samsung where they were only paying for
working dies on 8N, both from a functional unit standpoint and from a parametric standpoint, potentially making their overall cost even lower than the table below, which likely assumes that the customer pays for the wafer, not just the good dies.
Given that AD102 and GA102 have nearly the same die size, a rough guess as to the ratio of costs between AD102 and GA102 would just be the ratio of their wafer costs, which is glossing over even more potential cost savings from 8N due to only paying for 'good' dies as mentioned above:
Looking at a ~2.8x cost increase per die, and even that might be charitable, given that Samsung 8N costs have likely gone down considerably since that chart was produced, or at the very least, gone down more than TSMC 5nm class nodes have, which are still hugely in demand.
Obviously the manufacturing cost of the GPU die is only one
part of the entire BOM, but still likely the single largest line item.
As
@DegustatoR mentions, there's more to this here than just corporate greed, making giant chips on bleeding edge nodes isn't cheap.