GH200 NVL2 or 8x RTX 6000 Blackwell for running Kimi K2.6 / DeepSeek V4 locally? (5 devs, agentic coding)

In three linesDeveloper seeking optimal infrastructure (~$100-150k) to self-host Kimi K2.6 and DeepSeek V4 locally for 5-person team (agentic coding). Compares dual GH200 NVL2 (1.2TB unified memory, $95k) vs 8x RTX 6000 Blackwell (768GB VRAM, $140k). Single GH200 test: 23 tok/s decode at 2-bit quant, but slow prefill and models overflow into slower unified memory.

Summary generated by Claude — human-verified