First hardware for a self-hosted LLM: you almost certainly don't need a cluster

SAMPLE POST — replace body with the real article before launch.

Models with reliable tool calling already fit on a single card.

What you actually need

• One workstation-class GPU
• A 2GB embedder
• An OpenAI-compatible API

def capacity(input_tokens: int, output_tokens: int) -> float:
    return input_tokens / (input_tokens + output_tokens)

Limits of applicability

All numbers here are N=1: one card, one production, one stack.