Self-hosted LLM for 25 developers: prefix cache beats raw model choice

vLLM on Blackwell: 96.9% prefix cache hit rate, 452:1 input/output, and why teams keep inference in-perimeter for control—not token savings.