Chiplets are an intriguing idea. Theoretically you could move ROPs, memory controllers and cache off to a separate on-package die like AMD did with Zen 2. You would also need to move the "front end" that talks to the CPU. It'll be slower and use more power than a big fat chip though.
Why should small CPU dies with single digit performance increases cost more than GPUs 3-4x the size that substantially improve performance each generation ?
Chiplet designs are early days. It doesn't have to be slower at all. I think AMD has sort of proven that with Zen2. It's faster generation to generation on a magnitude we haven't seen in a long long time.
In GPUs yes the challenges are different. Of course both AMD and INtel are planning chiplet versions of their GPUs cause they will be slapping them on their CPUs. Intel is also potentially using their 3D Foveros chiplet stacking tech. Which if they pull it off is for sure better then what AMD is currently cooking and really will no different then one big hunk of silicon. The chiplets would be stacked and connected with traces which would actually be shorter then what they are currently using connecting one CCX to a second one.
As far as multiple chiplet GPUs go... I agree that is a different ball of wax. But I think all the players are looking at ways to make it happen. NV has already built chiplet research chips for AI stuff. Which are sharing massive amounts of data. AMD is no doubt at the least going to have a navi chiplet to slap on the console parts. Presumably their controller chip is going to have to address the GPU as well. Intel is a bit of an unknown but based on the little bits they have released. I think its a safe bet to assume they have some form of chiplet as well. Their integrated situation will greatly improve if they really do have a next gen using 3D stacked CPU and XE graphics. I suspect their cards are also going to be chiplet based somehow. Perhaps it will be as simple as higher end cards having an extra bit intended to be used for things like RTX.
As for the cost of GPUs... I don't know your right a lot of people seem ok with spending north of a grand on a GPU. I guess I'm old school but I remember when I could actually afford a 3DFX card and a second one for SLI didn't cost more then a server chip. We are literally at a point where you can pick up an Epyc 7352 (7nm 24 core 48 thread) server processor for less then the cost of the highest end consumer GPU. I consider that insanity myself. It basically means I could build a 24 core 128gb solid mid range server... for the cost of a decent gaming PC with a 2080ti in it. Seems stupid to me. But who am I to judge people willing to spend that money. But I mean I do also laugh at yuppies (do we still call em that right?) driving around in their expensive little german cars and more often stupidly oversized and over priced minivans (SUVs I think the kids call em).