the big reason I upgrade GPUs these days is for more VRAM for LLMs and diffusion models. I don't care (or need to care, really) as much about gaming -- along with great Proton support, running things from a midrange Linux-based gaming PC I have shoved in my home server rack works great via Steam's Remote Play (NoMachine also pretty good), but I play strategy/spreadsheet games, not twitchy FPS games.
my most recent upgrade was for a 4090, but that gives me only 24GB VRAM, and it's too expensive to justify buying two of them. I also have an antique kepler datacenter GPU, but Nvidia cut driver support a long while ago, making software quite a pain to get sorted. there's a nonzero chance I will wind up importing a Moore Threads GPU for next purchase; Nvidia's just way too expensive, and I don't need blazing fast speeds given most of my workloads run well inside the time I'm sleeping, but I can't be running at the speed of CPU; I need everything to fit into VRAM. I'd alternately be stoked for Intel to cater to me. $1500, 48GB+ VRAM, good pytorch support; make it happen, somebody.