It all depends on the VRAM configurations.
I can play all the games I like on an APU: 10+ years old, indy and online low requirements ones like TF2.
But if low powered SKUs with high VRAM are available I will probably buy 4 to 8 of these for playing with larger LLMs.
Card speed is irrelevant if the datasets and models don’t fit.
If VRAM is stingy, my money goes to Apple for a Mac Studio with M4 Ultra and 512GB unified memory.
The ball is in Nvidia’s court. Take my money or another company will get it.
AMDs software i do monitor. If they are good enough for LLM use on Linux I might consider them.
But Apple solution will be way more polished and supported, so AMD have to either improve their stack to rock solid level or be significantly cheaper (2x or more) to compensate for the time waste and frustrations.