You don't need that much for simpler models. Some can even run on a PI.

Qwen3 and Gemma models are fairly capable, they are slow-ish (a few tokens per second) but will run.

You can start building with cheap hardware and simple models, and use something more capable once you're more confident on the use case.