I think you might be influenced by the marketing of "the performance is more, so it should cost more" fallacy. The more pragmatic perspective should be, "what does it cost for Nvidia to produce those chips with the current state of technology". Basically, if we look at history as guide, the cost to produce the top chips mostly stay the same taking inflation into account.
Eg:
980Ti GPU production cost should be close to 1080Ti GPU cost
1080Ti GPU production cost should be clsoe to 2080Ti GPU cost
And with each new generation came a performance improvement for the same price due to advancements in logic design and process improvement. The logic wasn't, "oh the 2080Ti is 35% faster than the 1080Ti so it should cost 30% more". That's not how the narrative worked and if anyone starts to follow that logic then the battle is already lost.
I'm not saying what I think pricing
should be, but rather what it probably
will be. The simple fact is that at launch, every new generation of GPU tends to be in high demand, and there will be some people willing to pay that price. So Nvidia, AMD, and their AIB partners will take advantage of that.
You're also completely overlooking the increasing costs of technology. 980 Ti had a 601mm^2 die made on a very mature TSMC 28nm node, with 6GB VRAM. 1080 Ti had a 471mm^2 die on a newer TSMC 16nm node, so the shrink in die size helped offset the added price of the chips; it also moved to 11GB. 2080 Ti moved to faster GDDR5X memory, but also went to a massive 754mm^2 die size on TSMC 12nm — a minor respin of 16nm. That was a big jump, but where things really started to cost a lot more was with the sub-10nm nodes.
RTX 30-series is made on Samsung "8nm" (respin of 10nm), moved to GDDR6X on the top GPUs, and memory on the 3090 was 24GB (more than double the 2080 Ti). The die size is 628mm^2. Big chip, more expensive node, price more or less stayed the same on MSRP, but perfect storm of Covid and crypto ruined those prices.
Now RTX 40-series is moving to a cutting-edge TSMC "4nm" node (5nm respin). It's not the absolute bleeding edge of TSMC N3, which isn't quite ready I think, even for Apple, but it's definitely a lot more costly than Samsung 8N. How big will AD102 be? If it's roughly the same 600mm^2 as GA102, price of the chips basically just doubled — not counting VRAM or anything else. If power delivery and cooling need to be able to cope with 450W or even 600W, that costs more money as well.
The general rule of thumb is that a straight 50% increase in bill of materials also means there will be at least a 50% increase in retail pricing — usually more. And as I've said before elsewhere, everyone needs a profitable slice of the pie to keep their businesses running in the black. That means about 15% for retailer, 15% for distributor, 15% for AIB partner, and 15% for AMD/Nvidia/Intel.
If a card costs $250 in raw materials to make, which is probably a reasonable estimate for something like an RTX 3070, it needs to sell for at least $437 (give or take a bit). So if raw BOM on RTX 4090 is $500, minimum retail is $875, and being the halo card that basically gets doubled (Apple pricing). RTX 4080 might only cost $450 to make and still sell for under $800 just because it's intended to be more competitive.