In July 2026, China's information communication and AI computing infrastructure construction reached a historic milestone. The Shuguang 8000 (Dengfeng), China's first fully domestic 100,000-card AI supercluster built by Sugon, made its global public debut and was selected as a "Treasure of the Hall" at the 2026 World Artificial Intelligence Conference (WAIC). This is not just an exhibition of a super-large-scale computing system, but a landmark leap for China's AI infrastructure from the 10,000-card level to the 100,000-card level.

Computing Power Leap from 10,000 Cards to 100,000 Cards

On July 10, 2026, at the Photosynthesis Organization 2026 Intelligent Computing Application Conference, Sugon announced the official completion of the Shuguang 8000 (Dengfeng), China's first fully domestic 100,000-card AI supercluster, which was simultaneously connected to the National Supercomputing Internet. Just one week later, the system made its global debut in its physical form at WAIC 2026 and was selected as a "Treasure of the Hall" of the conference.

"100,000 cards" refers to the deployment of 100,000 or more AI accelerator cards within a single cluster. Compared to 10,000-card systems, a 100,000-card deployment tests not only the number of computing cards and theoretical peak performance, but also system architecture, network interconnection, memory access efficiency, energy efficiency control, and robust ecosystem application capabilities. The transition from "10,000 cards" to "100,000 cards" is not merely a linear increase in scale, but a systematic leap in foundational capabilities—marking the beginning of China's AI infrastructure construction moving from the 10,000-card level to the 100,000-card deployment stage.

Three Core Technological Breakthroughs: Redefining Ultra-Large-Scale Computing Power

"Super-Intelligence Fusion" Architecture: One System for All Scenarios

For a long time, ultra-large-scale computing clusters have faced a "curse": the larger the scale, the narrower the scenario adaptability—either specializing in high-precision scientific computing or focusing solely on low-precision AI training.

The answer provided by the Shuguang 8000 is "native super-intelligence fusion." The system abandons traditional partitioning methods, achieving native integrated fusion of all types of computing. From FP64 double-precision scientific computing to INT8 low-precision AI inference, full precision coverage is achieved. Researchers simulating protein folding, engineers running industrial simulations, and AI teams training large models—all scenarios are handled by a single base system.

This solves the pain point of needing cross-system data scheduling for multiple task types in the past, which involved cumbersome processes, and seamlessly covers diverse scenarios such as basic scientific research, large model training, AI inference, and industrial simulation.

Full-Chain Self-Development: Unblocking the "Governor and Conception Vessels" for Stable 100,000-Card Operation

The stable operation of a 100,000-card cluster is a systematic engineering challenge; any weak link in the chain can trigger a "butterfly effect." The Shuguang 8000 achieves full-chain independent controllability from the underlying chips to upper-layer application services:

Chip Layer: Supported by domestic chips such as Hygon, with general-purpose processors, AI accelerator chips, and switching network cards all domestically developed.

Network Layer: Self-developed scaleFabric, a native RDMA high-speed network similar to InfiniBand, achieving "zero packet loss and millisecond-level fault recovery" at the 100,000-card scale.

Storage Layer: ParaStor distributed storage supports massive data reads and writes in large model training and scientific computing, ranking first in both the production full-node and 10-node performance categories on the 2026 Global IO500 List.

Cooling Layer: Immersion phase-change liquid cooling technology supports MW-level high power density deployment per cabinet, with a PUE as low as 1.04, enabling green and efficient computing.

Scheduling Layer: The Gridview 7.0 intelligent computing scheduling platform adopts a MetaStack+K8s dual-fusion architecture, unifying the scheduling of supercomputing and AI computing power.

Structural Innovation: The world's first high-density cabinet structure increases the computing density of a single computing unit by 20 times compared to other super nodes.

These seemingly disparate technologies collectively form the complete "skeleton" of an ultra-large-scale computing system.

Real-Machine Combat: A "National Treasure" Validated by Over 300 Applications

The Shuguang 8000 is by no means a "showpiece" in the lab. Currently, the system has completed optimization for over 300 super-intelligence fusion applications, covering more than 20 scientific research and industrial fields, including large models, robotics, innovative drugs, new materials, quantum computing, astronomy and meteorology, and automotive. Among these, over 70 applications have achieved 10,000-card scale expansion.

In the scientific research field, it supports cutting-edge scientific computing such as protein folding simulation, trillion-atom-level water molecule dynamics simulation, and quadrillion-grid ultra-large-scale turbulence simulation. In the industrial field, it serves AI applications such as large model training, scientific intelligent agents, and industry intelligent agents for continuous operation.

From "Built Successfully" to "Well-Used and Replicable"

The completion and debut of the Shuguang 8000 carry multiple profound significances:

Establishing Standards for 100,000-Card Domestic Computing Power Construction. It provides a replicable, fully domestic construction model for 100,000-card intelligent computing infrastructure, shifting the evaluation focus of large-scale computing centers from scale stacking to more comprehensive system capabilities.

Connecting to the National Computing Network, Serving the Whole Country. The system has been integrated into the national integrated computing network via the National Supercomputing Internet, offering open, inclusive, efficient, and convenient computing services to research institutions, universities, enterprises, and individual users.

Initiating the Construction of a Second System. Sugon has entered into a strategic cooperation with the Beijing Academy of Artificial Intelligence (BAAI) to start the development and construction of a second fully domestic 100,000-card super-intelligence fusion computing system, marking the transition of 100,000-card AI infrastructure from a landmark project to continuous deployment.

From a single chip to a complete system, and then to computing power serving thousands of industries, competition in the AI era increasingly reflects long-term accumulation and comprehensive capabilities. The completion and debut of the Shuguang 8000 (Dengfeng) reflect that China's AI infrastructure is moving from "single-point technology competition" to a new stage of "system capability competition."

Tan Guangming, an academician of the Chinese Academy of Engineering and a researcher at the Institute of Computing Technology, Chinese Academy of Sciences, pointed out that the 100,000-card AI infrastructure is beginning to transition from a landmark project to continuous deployment, which will further provide new infrastructure support for the construction of the national integrated computing network, breakthroughs in cutting-edge science, and the intelligent upgrading of industries.

As artificial intelligence moves from technological exploration to large-scale application, computing infrastructure is stepping from behind the scenes to the forefront, becoming an important yardstick for measuring the level of AI development. The debut of the Shuguang 8000 marks that China's AI computing infrastructure construction has officially entered a new era of "full precision, 100,000 cards, and full domestic production."