What does on-premise LLM deployment actually cost and what hardware is sufficient?

We have a client who cannot send data to external APIs under any circumstances. We need to run a capable model entirely on-premise. The workload is moderate — maybe 50 concurrent users, document Q&A and summarization. What GPU/CPU configuration is realistic for running a 70B-class model at acceptable latency, and what are the operational costs teams have seen vs. the API alternative?