Inspur Releases CPU-Native Liquid Cooling Rack and SD200 Super Node, Reshaping Agent Infrastructure

At the 2026 Open Computing Conference, Inspur Information launched the industry's first CPU-native liquid cooling rack server and an upgraded version of the YuanBrain SD200 super node AI server, aiming to address two core challenges: large-scale Agent operation and multi-model collaboration.

Background: The Agent Era Imposes New Demands on Infrastructure

According to IDC data, the enterprise market size for AI Agents in China reached approximately 19 billion yuan in 2025, with a projected compound annual growth rate of over 110% from 2025 to 2028. Gartner predicts that by 2026, 40% of enterprise applications will integrate task-based AI Agents. Agent applications require task decomposition, tool invocation, and multi-round collaboration, placing new demands on computing power: CPU importance has significantly increased (research shows CPU processing time accounts for up to 90.6% of end-to-end latency), and rack power density is soaring (domestically reaching 300 kW within the year, with some globally entering the megawatt level). Traditional air cooling has an upper limit of only 40-50 kW, making liquid cooling a necessity.

Key Products: CPU-Native Liquid Cooling Rack and SD200 Super Node

• CPU-native liquid cooling rack server: A single rack supports up to 384 CPUs (compatible with x86 and ARM), enabling over 40,000 Agents to run collaboratively—40 times the capacity of the previous "Qianqixia" solution. It adopts a native liquid cooling architecture, covering all heat-generating components such as memory, network cards, and optical modules, achieving zero hoses, zero cables, and zero fans. It supports hot maintenance, improving O&M efficiency by over 100%. Single-rack power consumption can reach up to megawatt levels, adapting to the trend of 800V high-voltage power supply into racks.
• YuanBrain SD200 super node AI server: After the upgrade, single-token generation time has dropped to 4.77ms (the first in China to break 5ms), with first-token latency reduced by 35%. It can simultaneously deploy four trillion-parameter large models and has been adapted to mainstream open-source models such as Kimi K2.6, DeepSeek V4, GLM 5.2, and MiniMax M3. An enterprise edition (16-card Scale Up) has also been launched, reducing first-token latency by over 40%, lowering the deployment threshold for small and medium-sized enterprises.

Multi-Model Fusion API: Enhancing Agent Intelligence

Inspur has launched a multi-model fusion API on the YuanBrain EP AI platform, which distributes the same task to multiple candidate models simultaneously. A review fusion model compares consensus, divergence, omissions, and unique perspectives before outputting a unified result. It achieved a score of 53.9% on the DRACO test, higher than any single model. Simple tasks are handled by a single model, while complex long-chain tasks invoke multiple models, avoiding resource waste.

Impact and Trends

The competitive focus of Agent infrastructure is shifting from single-model support to system-level collaboration. CPUs, GPUs, and software platforms are working more closely together: CPUs handle Agent instances and tool invocation, GPUs handle inference and token generation, and software platforms handle task orchestration and result fusion. This year, top domestic internet companies have directed almost all new CPU server procurement toward Agent businesses. Inspur's release marks a new phase for Agent infrastructure, characterized by large scale, high density, and liquid cooling.