You don't need a bunch of a100s. The models you can download and run locally on a consumer GPU or Apple Silicon will say whatever you tell them to. You won't be getting GPT-4 level output (with some exceptions), but it's close.