Inspur Unveils CPU-Native Liquid Cooling Rack and Yuanbrain SD200 Enterprise Edition, Supporting 40,000+ Agents Per Cabinet

The core mission of AI infrastructure is rapidly shifting from supporting large language model inference to enabling large-scale deployment of intelligent agents. Inspur addressed this shift at its 2026 Open Computing Conference with two major product announcements targeting both agent scalability and multi-model collaboration.

Agent workloads differ fundamentally from simple Q&A — they require task decomposition, tool calling, multi-turn collaboration, and continuous operation, demanding a complete redesign of underlying computing infrastructure. According to IDC, China’s enterprise AI Agent market reached approximately 19 billion RMB in 2025, with a projected compound annual growth rate exceeding 110% from 2025 to 2028. Gartner projects that 40% of enterprise applications will integrate task-oriented AI agents by 2026.

To meet these demands, Inspur unveiled the industry’s first CPU-native liquid cooling rack server. Each cabinet supports up to 384 CPUs based on the open OCM architecture, compatible with both x86 and ARM, and can sustain over 40,000 agents running concurrently — 40 times the capacity of Inspur’s QiQianXia solution from April 2026. The OCM architecture also allows cross-generational and cross-architecture processor compatibility, significantly compressing development cycles.

On the cooling front, Inspur introduced a native liquid cooling approach that abandons traditional air-liquid hybrid designs. All heat-generating components — including memory, network cards, optical modules, and SSDs — are integrated into a single liquid cooling system. Each 2U ultra-thin node packs 16 CPUs laid flat on a unified cold plate, achieving a cable-free cabinet design that improves maintenance efficiency by over 100%.

To make agents smarter, Inspur launched a multi-modal fusion API on its Yuanbrain EPIA platform. The system distributes the same task to multiple candidate models, then uses a review model to compare consensus, disagreements, and omissions before producing a unified output. The approach scored 53.9% on the DRACO benchmark, outperforming any single model in the test pool.

Inspur also released the Yuanbrain SD200 supernode enterprise edition. The previous SD200 could already deploy four trillion-parameter models simultaneously with 8.9-millisecond token generation. The 2026 edition pushes this to 4.77 milliseconds — the first Chinese solution to break the 5-millisecond barrier — while reducing first-token latency by 35%. The enterprise edition scales down to 16 GPUs from 64, lowering first-token latency by over 40% and enabling companies that previously could only deploy hundred-billion-parameter models to bring trillion-parameter models into production.

The SD200 now supports mainstream open-source models including Kimi K2.6, DeepSeek V4, GLM 5.2, and MiniMax M3, with the multi-modal fusion API available for general access.