(Yicai) July 15 -- Huawei Technologies will debut the Chinese tech giant's flagship supernode, the Atlas 950 SuperPoD, at the 2026 World Artificial Intelligence Conference and High-Level Meeting on Global AI Governance, which opens at the end of this week.

The Atlas 950 SuperPoD capabilities include the industry's largest 1024-card supernode with 1 EFLOPS FP8 computing power, the largest 256-terabyte global single virtual address space, terabyte-level neural processing unit interconnection bandwidth for training and inference performance, and an ultra-low round-trip time latency of three microseconds, Huawei revealed on its website on July 13.

The WAIC will run from July 17 through 20 in Shanghai. While the main focus in recent years was on large language models and AI chips, a distinct signal conveyed for this year's conference is that the global AI competition is rapidly shifting its focus toward infrastructure, which involves organizing an increasing number of AI chips into an efficient and collaborative computing system.

Supernodes integrate hundreds to thousands of AI chips into a single large computing power pool through high-speed interconnection tech.

The Ascend supernode computing products have been successfully implemented at scale in over 20 industry scenarios, providing low-latency and high-throughput computing services for internet companies, LLM vendors, and innovative institutions in research, education, and healthcare, Huawei pointed out.

Other Chinese companies will also show their achievements in organizing AI chips into an efficient collaborative computing system at the upcoming WAIC. For example, telecom equipment and consumer electronics maker ZTE will present its high-performance Matrix supernodes, developed in collaboration with several domestic graphics processing unit firms.

In addition, multiple Chinese GPU companies will release their supernodes that they have primarily developed.

Although different manufacturers have varying technical approaches, their supernodes share a highly consistent goal of enhancing the resource utilization efficiency of the entire AI system through greater communication bandwidth, lower network latency, and unified memory addressing to support the efficient operation of AI models.

The performance of a single Ascend chip is only about one-third that of Nvidia's Blackwell architecture GPU, according to industry research firm SemiAnalysis. However, through the use of supernodes and clusters, the overall performance can be effectively enhanced, compensating for the computing power shortcomings of Chinese chips.

"Supernode and cluster technologies can continuously provide ample computing power for China's AI industry," Eric Xu, rotating chairman of Huawei, previously told Yicai. Although a single Ascend chip still lags compared to Nvidia, and there are also differences in the ecosystem, after forming supernodes and clusters, the Shenzhen-based company is confident in the competitiveness of its products, he stressed.

Editors: Tang Shihua, Martin Kadiev