Summary of Key Points
At the 2026 World Artificial Intelligence Conference (WAIC), Huawei showcased the Ascend 950 supernode for large-scale AI training and inference for the first time, marking a shift in the competition for AI infrastructure from a focus on "single-chip performance" to "system-level coordination capabilities." The Ascend 950 utilizes proprietary technology to achieve efficient interconnection of thousands of chips, boasting outstanding computing power and coordination abilities. Additionally, a fan-cooled version was introduced to lower the barriers for enterprises to adopt this technology. Huawei's Ascend supernodes have been commercially adopted in various industries, and there are signs of breakthroughs in overseas markets as well. Companies such as Baidu and Alibaba have also unveiled their own supernode products, indicating that the market for domestically produced supernodes is expected to grow rapidly, with domestic chips likely taking a dominant position.
Detailed Analysis
Why Have SuperNodes Suddenly Become So Popular in the AI Community?
Previously, AI competition mainly revolved around the computing power of individual chips—whichever chip could perform the most calculations per second. However, AI models have become increasingly complex, with parameters increasing from tens of billions to trillions and requiring the processing of longer dialogues as well as real-time responses from intelligent agents (such as AI assistants). This necessitates the use of hundreds or even thousands of chips working together. The challenge then becomes how to rapidly transfer data between these chips, share memory, and coordinate their tasks efficiently—this is where "system-level capabilities" come into play. Supernodes are designed to solve this issue by organizing multiple chips into a highly efficient "supercomputer cabinet," allowing them to function as a unified entity, thus becoming the core of current AI infrastructure competition.
What Makes the Huawei Ascend 950 Supernode So Advanced?
In simple terms, the Ascend 950 is an advanced system that enables efficient collaboration among thousands of chips:
- Powerful Computing: A single cabinet can accommodate up to 64 Ascend chips, and when connected in maximum capacity, it can process 1024 Ascend chips, providing computing power equivalent to 100 trillion floating-point operations (FP8) or 200 trillion FP4 operations per second—more than sufficient for training large models like ChatGPT.
- Fast Collaboration: Using Huawei's proprietary "Lingqu Interconnection Protocol," the data transfer bandwidth between chips reaches terabits per second, allowing for the transmission of several high-definition movies in a single second with virtually no latency (just 3 microseconds).
- Unified Memory: With 256TB of global memory addressing, all chips can access the same large memory pool directly, eliminating the need for data copying, which significantly improves efficiency during model training.
These combined features address the issues associated with slow model training and inference.
How Does Huawei Help Enterprises Save Time and Money When Using AI?
In addition to the powerful Ascend 950, Huawei also introduced the Atlas 850E fan-cooled supernode:
- While traditional AI supernodes often use liquid cooling systems that require specialized data centers, the Atlas 850E utilizes Huawei's VCE phase change cooling technology, which can be installed in standard fan-cooled facilities without any infrastructure modifications.
- A single fan-cooled supernode can be expanded to accommodate up to 96 chips, enabling enterprises to utilize large-scale AI computing power at a lower cost, making it more accessible to smaller and medium-sized businesses.
How Well Have Ascend Supernodes Been Implemented So Far?
- Domestic Adoption: The Ascend 384 supernode, released in 2025, has sold over 750 units and been applied in industries such as the internet, telecommunications, finance, and healthcare. It is also the only domestic supernode capable of training industry-leading models (SOTA). At the conference, Huawei demonstrated cases from over 20 sectors, including financial risk management and medical image analysis.
- Overseas Progress: There are reports that Korean cloud providers have ordered 2,000 Ascend 950 chips, and discussions with Russian and Malaysian companies are underway (although not officially confirmed), indicating demand for domestic supernodes in overseas markets.
The Current Competitive Landscape and Future Trends in the Supernode Market:
- Increasing Competition: Besides Huawei, other players such as Baidu (Tianchi 512 supernode), Alibaba (Zhenwu M890x Panjiu AL128 supernode), and ZTE (jointly developed with Bitwise and Mujixi) have also showcased their supernode products, intensifying the competition.
- Growing Market Size: Huatai Securities predicts that the market for domestic supernodes could reach 341.4 billion yuan by 2028, indicating rapid annual growth.
- Domestic Dominance: Jibang Research suggests that in 2026, Chinese manufacturers like Huawei and Cambricon will account for nearly 80% of the Chinese AI server chip market, while foreign companies (such as NVIDIA) may see their share decline. This indicates that domestic AI infrastructure is taking the lead in replacing imports.
Conclusion
AI competition has evolved from a focus on individual chips to a emphasis on system-level coordination, with supernodes playing a crucial role. The Huawei Ascend 950 excels in both performance and accessibility, and it has been successfully implemented in various industries overseas. The market for domestic supernodes is expected to expand significantly, with domestic chips becoming the standard. This development is vital for the autonomy and control of China's AI industry.