The other problem is with latency: Deepseek 34B on A100s seem slower than GPT-4 but perhaps it will be better on H100s.