China’s First Fully Domestic 100,000-Card AI Supercluster Launched in Zhengzhou, Integrated Into National Supercomputing Internet

The 2026 Intelligent Computing Application Conference hosted by Photosynthetic Organisation opened in Zhengzhou on 10 July. At the event, Sugon unveiled Shuguang 8000 (Dengfeng), the country’s first fully domestic AI supercluster with a capacity of 100,000 compute cards, which has been connected to the National Supercomputing Internet upon launch. China’s artificial intelligence computing infrastructure has crossed the 10,000-card threshold and entered a new phase of large-scale deployment at the 100,000-card tier.

Massive demand for computing power continues to expand amid widespread uptake of large model training, AI for Science research and autonomous agent applications. Computing workloads have grown more complex, featuring high concurrency, mixed precision and multi-task parallel processing, rendering clusters built around 1,000 or 10,000 cards unable to match industrial operational requirements. 

Industry practitioners confirm that 100,000-card supercomputing clusters have moved beyond conceptual demonstration and become a baseline infrastructure capability for next-generation AI systems. 

Deploying a 100,000-card system poses far broader technical challenges than simply stacking hardware to boost theoretical peak performance, placing rigorous demands on integrated capabilities spanning system architecture, high-speed interconnection, mass data access, energy efficiency control and cross-industry ecological compatibility.

Shuguang 8000 (Dengfeng) delivers a replicable fully domestic construction template for large-scale intelligent computing infrastructure at the 100,000-card tier. Built on a native hybrid supercomputing-intelligent computing architecture, the system abandons segmented resource partitioning frameworks to natively unify high-precision scientific simulation and low-precision AI computation. 

It delivers full-spectrum precision support ranging from FP64 double-precision floating-point arithmetic down to INT8 integer operations, delivering seamless coverage for fundamental scientific research, large foundation model training, AI inference and industrial digital simulation.

Sugon’s self-developed scaleFabric InfiniBand-alike native RDMA high-speed network maintains stable, lossless connectivity across the full 100,000-card cluster. The ParaStor distributed storage platform handles extreme-volume data read-write cycles generated during large model iteration and scientific simulation workflows. 

Backed by an open collaborative computing ecosystem and standardised cloud computing service frameworks, the platform delivers end-to-end technical support covering underlying hardware, scheduling middleware and industry-specific application adaptation for academic institutions, industrial operators and algorithm developers.

Integrated into the nationwide unified computing network via the National Supercomputing Internet, Shuguang 8000 (Dengfeng) opens affordable, high-efficiency and accessible computing resources to university research teams, commercial enterprises and individual developers. 

Over 300 hybrid supercomputing-intelligent computing applications have undergone targeted optimisation on the cluster’s core nodes, covering more than twenty vertical sectors including large language models, biomedicine, advanced new materials, quantum computing, meteorology and astronomical simulation. 

More than 70 of these applications support scaled execution at the 10,000-card level, with landmark scientific workloads including protein folding simulation and trillion-atom molecular dynamics successfully deployed to validate stable performance and industrial transformation value under extreme high-load operating conditions.

A strategic cooperation framework between Sugon and the Beijing Institute of AI for Science was signed concurrently with the conference opening, initiating development and construction of a second fully domestic 100,000-card hybrid supercomputing-intelligent computing cluster. 

Large-scale full-precision computing centres built to the 100,000-card standard will transition from one-off demonstration projects to widely replicable industrial assets. As the global artificial intelligence sector shifts toward rational, industry-focused development, such large-scale domestic computing infrastructure will become standardised foundational hardware to support full rollout of AI-enabled industrial workflows and accelerate original scientific discovery and digital economic expansion.