China deployed first 100,000-GPU cluster, running 300+ production applications on all-domestic accelerator stack.
On July 10 in Zhengzhou, Sugon announced the completion of its fully domestically-manufactured Sugon 8000 (Dengfeng) AI supercluster, which will simultaneously be integrated into the National Supercomputing Internet and China's Integrated Computing Power Network.
This is not simply an oversized 10,000-card cluster. The leap from 10,000 to 100,000 cards is not a matter of adding more units—it represents an entirely different engineering challenge.
The bottlenecks are numerous: maintaining network throughput without congestion, preventing storage systems from becoming unstable, managing heat dissipation without failure, and ensuring scheduling doesn't become chaotic. With each order-of-magnitude increase in scale, system complexity grows exponentially.
A common problem exists across current AI computing centers: clusters built for large language model training cannot handle scientific computing, and those designed for scientific computation cannot train trillion-parameter models.
Sugon 8000 addresses this with "native super-intelligence fusion"—a single unified system capable of both handling FP64 double-precision scientific computing tasks and sustaining trillion-parameter large language model training.
While the "super-intelligence fusion" concept is not new, Sugon 8000 stands as the industry's first 100,000-card-scale computing cluster built on a native super-intelligence fusion architecture and supporting full-precision computation. It establishes a practical pathway for optimizing domestic computing resource allocation—by constructing a "super-intelligence fusion" platform that pools distributed supercomputing and AI computing resources, enabling unified scheduling across platforms and architectures. This approach has significant value for improving domestic computing utilization rates and reducing redundant infrastructure investments.
The development timeline was: system design in 2024, engineering construction completed in 2025, 30,000 cards operational in February, an additional 60,000-card AI4S cluster added in April, and the full 100,000-card deployment completed in July.
According to Li Bin, Senior Vice President of Sugon, the Sugon 8000 demonstrates end-to-end domestically-developed capabilities across chip design, computing, storage, networking, cooling, management, and service systems.
ParaStor distributed storage supports both large language model training and scientific computing involving massive data I/O, ranking first in both the "production full-node" and "10-node" categories on the 2026 global IO500 rankings.
Leveraging Sugon's liquid cooling technological advantages, the system effectively supports MW-level high power density deployment. Through domestically-developed coolant and year-round passive cooling methods, Sugon 8000 achieves significantly improved energy efficiency.
Sugon has simultaneously integrated Sugon 8000 into the National Supercomputing Internet, offering computing services to research institutions and industry. The system has completed over 300 application adaptations across more than 20 fields including materials science, electromagnetics, quantum computing, biomedical research, astronomy, and meteorology, accumulating over 70 large-scale 100,000-card computing tests.
Particularly noteworthy is Sugon 8000's adoption of an open AI computing architecture that supports multiple brands of AI accelerators and is compatible with mainstream ecosystems—developers and researchers can leverage existing tool chains without relearning proprietary systems, and existing models and applications can migrate quickly online.
During the conference, Sugon established a strategic partnership with the Beijing Institute for Scientific Intelligence, formally launching development and construction of a second fully domestically-manufactured 100,000-card super-intelligence fusion computing system.
Eastern data transfer to western regions, China's Integrated Computing Power Network, and the National Supercomputing Internet all share one objective at the policy level: enabling computing power to flow as fluidly as electricity.