They released a few good value GPUs for LLM inference about a year ago: more memory than AMD and NVIDIA consumer GPUs, not too expensive, but also not great tokens/watt.

I am not sure whether you can find those in stock anywhere.