Exactly. That's as if fast food chain just before serving food to customer added ingredient that would made food taste worse (and or be less healthy) in order to be able to sell you same food without this ingredient at premium price.
We don't see such behavior because in fast food industry we have heavy competition. Not so much in GPU market where two companies have most of the market and entry barriers are high.
I don't know, as I'm familiar with it early in the development of a process node they sell every top-bin-SKU they can make, and as the process matures they just have too many top-bin parts. I can only guess your argument is that in a competitive market they would start dropping the price of the top-bin SKU to move them, but that reduces the overall profit of the venture which may drive them to increase initial pricing, if profit margins are thin enough.
Basically what I'd argue is that the pricing is the symptom, not the down-binning. Down-binning exists to maintain & stabilize SKU distinction and pricing, as yields change over the life of a product.
If that was what the market wanted, don't you think that is what the market would get? I think the trouble is you are thinking "the marginal cost on a mature process" would be a few dollars.
You're right, the silicon would be a few dollars- but after testing and packaging and all that jazz, I believe most mainstream high performance desktop CPUs have a marginal cost of around $30. (Why not cheaper, like a 10MHz ARM chip? Package is expensive, due to cooling needs and pins for power & DRAM interface) So, perhaps you bring it to market for $40.
This chip you're selling probably performs like a mid-range part in today terms, in the $100-150 range. But when you consider TCO due to power draw and cooling, the numbers start to get closer.
I haven't carefully laid this out on paper or anything, but point being, if such a chip would sell so well, why is nobody doing it?
It becomes much clearer if you increase the contrast by comparing today's parts to those from a decade ago. The clocks are similar, but the performance has come a long ways due to improvements in IPC, multicore, updates to the memory interface, etc.
The extra value a producer is able to capture from price discrimination due to a relative lack of competition can be thought of as one of the ways in which the deal between producers and society called "intellectual property" functions to shift rewards for innovation toward to the producer.
Something thought "insufficient" for one end, might be in proper indeed for some other end.