LLM Throughput Capacity Analysis

I’m trying to estimate our system capacity for long-context LLM calls, based on our message workload.

Assumptions per message (per alert):
• Avg input tokens: 34K
• Avg output tokens: 5K

GPU setups to compare:
1. 8× RTX 6000 Ada (48GB) = 384GB total
2. 2× RTX 6000 Ada (48GB) = 96GB total
3. 4× H200 (141GB) = 564GB total

Models to compare on each setup:
• DeepSeek v3.2
• GPT-OSS 20B
• GPT-OSS 120B


So what I need is
- Who many messages does the model handles per day? And how did I got the numbers ?
- ⁠the messages will be calculated as 10 hours per day since it’s the working hours