Advancements in efficiency with regard to the manufacturing processes across the board cause manufacturing costs to DROP which more than makes up for inflation when it comes to tech.
That's only true for manufacturing on the same process and even then, only for when economy of scale applies. There is nothing inherent about VLSIC production, compared to other manufacturing forms, that automatically reduces costs over time. But even if it did, R&D, marketing, and shipping costs do increase - partly because of inflation, partly because designing multilayer, multiple billion transistor chips becoming increasing more challenging design, requiring more resources to do so.
But do note that I never said that inflation was the
sole reason behind the prices we see today - I even pointed out that when accounting for that, the likes of the RX 6900 XT was 37% higher than an inflated HD 5970.
GTX 295 (January 2009): $500
GTX 480 (March 2010): $500
GTX 580 (November 2010): $500
GTX 680 (March 2012): $500
GTX 780 (May 2013): $500
GTX 980 (September 2014): $550 <- Huh?
GTX 1080 (May 2016): $600 <- Huh?
RTX 2080 (September 2018): $700 <- WTH????
RTX 3080 (November 2020): $700
The GeForce GTX 780 was released with
an MSRP of $649, and don't forget that the
Founder's Edition of the GTX 1080 was $699 - the same applies for the 2080, where the
FE model was $799. So Nvidia has been churning high prices for longer that the list suggests.
It could be argued that the 480 and 580 were the same price, because they ostensibly used the same chip: the GF110 in the 580 is a minor layout alteration of the GF100 in the 480, to reduce power consumption. Both were manufactured on the same process, too.
The GK104 in the 680 wasn't a large chip, even by the standards of the time, at just 294 mm2 with 3.54b transistors - the GK110 in the 780, made on the same process at the GK104, is far larger: 561 mm2 with 7.08b transistors.
So the price hike over the 680 should be expected. The GM204 in the 980 contained substantially fewer transistors than the GK104, at 5.2b - so even though it used the same process, the chip was a lot smaller at 398 mm2. The GP104 in the 1080 was smaller again, at 314 mm2, but the complexity had increased once more, jumping back to just over 7b transistors.
By then, Nvidia were using a semi-custom process node with TSMC, and I doubt they would offer any price reduction for Nvidia to use it. A fully custom node was used for the TU104, in the 2080, and that was another big step up in complexity: 13.6b in a 545 mm2 die.
So when you add that information into the list, the changes in prices become a little more understandable:
GTX 295 - $500 - 470 mm2, 1.4b trans, TSMC '55nm'
GTX 480 - $500 - 529 mm2, 3.1b trans, TSMC '40nm'
GTX 580 - $500 - 520 mm2, 3.0b trans, TSMC '40nm'
GTX 680 - $500 - 294 mm2, 3.5b trans, TSMC '28HP'
GTX 780 - $649 - 561 mm2, 6.1b trans, TSMC '28HP'
GTX 980 - $550 - 398 mm2, 5.2b trans, TSMC '28HP'
GTX 1080 - $599/$649 - 314 mm2, 7.2b trans, TSMC '16FF'
RTX 2080 - $699/$799 - 545 mm2, 13.6b trans, TSMC '12FFN'
RTX 3080 - $699 - 628 mm2, 28.3b trans, Samsung '8NN'
So for the period of a constant $500, 2 chips were virtually the same, and one was a notable reduction in size. Nvidia would have had to absorb increases in their own costs, but they could be offset by the points just made. The real oddities, though, are the 980 and the 2080 - the former can be explained by the fact that it was on a very mature node, and also a lot smaller than its predecessor. The latter, specifically the FE version, was just ridiculously priced.
Now understandable isn't the same as accepting or condoning such prices, of course, because ultimately Nvidia were able to keep the price the same between the 295 and 480, over a 4 year period, despite moving to a notably larger and far denser chip, and on a new process node. But that was 11 years ago, a time when was heavily focused on improving their gross margins and it's interesting to note that in their
2010 financial report, they admit that average selling prices and competition play just as much of a factor in their gross margins, as do manufacturer costs/yields and the competition.
I get the feeling that Nvidia first tested the waters with the price of the GTX 780, and found that it wasn't ready for a $650+ not-quite-the-very-best graphics card. Cue the 980's drop, to sweeten the market ready for Pascal - which, as we known, sold extremely well. Corporates can only be as greedy as the market allows, and with the die set by Pascal, it was no surprise to see the 2080 and 3080 hitting $700.
That said, the GA102 is absolutely massive - only the GV100 and GA100 are larger (at over 800 mm2 apiece). The wafer yields can't be all that great, so although it's $200 more, I'd argue that the 3080 is back to the $500 days, in terms of what you're getting for your money. Or put another way, I find the price of that far more acceptable than the price of the FE 1080 and 2080.