Inspur MetaBrain SD200 Ultra Runs 2.8T Kimi K3 at 5.85ms/Token
Inspur MetaBrain SD200 Ultra Packs 128 Domestic AI Chips for 2.8T Kimi K3 Under 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.
At AICC2026, Inspur Information launched the MetaBrain SD200 Ultra, an AI supernode built from 128 tightly coupled domestic (Chinese-made) AI chips. The system provides 8 TB of unified accelerator memory across the cluster. Inspur claims sub-5.85 ms per token latency when serving the 2.8-trillion-parameter Kimi K3 model from Moonshot (Kimi). The launch signals growing domestic Chinese infrastructure capable of serving frontier-scale models without foreign accelerators.
Inspur MetaBrain SD200 Ultra Packs 128 Domestic AI Chips for 2.8T Kimi K3 Under 5.85ms/Token
At AICC2026, Inspur Information launched the MetaBrain SD200 Ultra supernode with 128 tightly coupled domestic AI chips, 8 TB of unified accelerator memory and a claimed sub-5.85 ms/token latency running 2.8-trillion-parameter Kimi K3.
