You have to specify which quantization you find acceptable, and which context size you require. I think the most affordable option to run large models locally is still getting multiple RTX3090 cards, and I guess you probably need 3 or 4 of those depending on quantization and context.