Also, my concern is not for data centers running creepy ML models that invade users' privacy, but for the average home user who wants to get started with ML and at the same time get decent frame rates when they play a game.
Getting started with ML doesn't require a GPU at all. Realistically, any video card or even a modern CPU is more than sufficient to get started with ML learning.
It's a myth that $1000 gaming video cards are necessary or even helpful for learning basic ML.
My argument is they should and could have scaled up the production rather than just shuffling around which group they hurt or help at the expense of which other (be it miners, ML folks or gamers). This isn't the 'right thing', it just helps the group you prefer on this round.
A sudden implosion of demand combined with a flooding of the used GPU marked. Better not plan on bringing out any new GPU architectures when your warehouses are still full of the previous generation.