Inspur MetaBrain SD200 Ultra Runs 2.8T Kimi K3 at 5.85ms/Token
The notable point is chip aggregation: 128 domestic accelerators tightly coupled into one supernode with 8 TB unified memory lets very large models run entirely on non-NVIDIA hardware, a direct alternative to NVL72-style designs. Hardware planners and inference teams evaluating sovereign or cost-sensitive deployments should benchmark it against NVIDIA's reference racks. Given Kimi's reach among Southeast Asian developers, domestic serving stacks could lower barriers to hosting such models regionally.