If you want to run LLM > \~50 GB, you could nearly don't bother about a GPU. I've run Deepseek Qwen3 thinking, some model around 200GB, with a PC with 192 GB ram AND a 5090, and before with a 3090. In such setup, the GGUF inference is so mostly in PC RAM that the GPU VRAM doesn't improve a lot the resulting performance (T/s)

So if you are ok waiting for 10-30 minutes a full answer (depending of the provided context and the size of the result), but want the best quality a LLM can give you, go for a 256GB DDR5 PC, with then the biggest GPU (VRAM) you can afford, which is at least ddr6.

If you want speed and some quality, some of the best scenario may be to have LOT of 'cheap' GPU and use GGUF inference, which can combine the VRAM of every GPU as if it was a single one. So if you take 4 GPU with 16 GB (like 5060ti or 5070ti) ram you will be just slightly little slower than two 5090 combined, but for maybe around 1/4th to 1/8th of the price depending of the current pricing. But depending of the scenario you may need to have good pci connexion else the pci transfer will slow down the inference and gpu will simply wait for data.

I'll soon try 5090 + 3090 + 5090 on a Thunderbolt 5 configuration, with 256gb ddr5 pc ram. Not sure what will be the resulting speed and exactly what will be the biggest LLM I'll be able to run, but I'll make myself that little surprise since it is my hobby for now 2-3 years...