(Science and Technology Innovation Board Daily) On July 18, it was reported that from July 17 to 20, the 2026 World Artificial Intelligence Conference and the high-level conference on global AI governance were held in Shanghai. The conference, themed “Intelligent Partners, Co-creating the Future,” was attended by official representatives from over 100 countries and international organizations, as well as guests from industry, academia, research institutions, and other sectors.
Meanwhile, the National Development and Reform Commission issued the (Action Plan for Promoting Cooperative Development of Artificial Intelligence), proposing various measures in eight areas, including high-quality data supply, intelligent computing power inclusive initiatives, and open-source ecosystem sharing, among others.
Computing power has long been the cornerstone of the development of the AI industry. Unlike last year, when computing power products were more concentrated in single computing chips, at this year’s WAIC, multiple major companies and enterprises in the computing power supply chain released results related to supernodes.
Among them, Huawei Ascend 950 supernode (Atlas 950 SuperPoD) was unveiled to the public in its full hardware form for the first time. It is aimed at compute demands for building large-scale data centers, training trillion-parameter models, and high-concurrency inference scenarios at the center. Based on the Lingqu interconnect protocol and supernode architecture, the product achieves the industry’s largest 1024-card scale, providing 1 EFLOPS FP8 and 2 EFLOPS FP4 blazing compute performance. It also features a unified 256TB global memory addressing space. Leveraging TB-level NPU interconnect ultra-high bandwidth and 3μs ultra-low RTT latency, it breaks through communication bottlenecks for cluster internal computing, storage, and other resources.
Alibaba Cloud officially released the Lingjun Zhenwu M890 supernode instance, which—using a supernode form factor—delivers a high-speed, plug-and-play 64-card compute unit. This is Alibaba Cloud’s first time to offer supernode-form AI compute services publicly through the cloud. The Lingjun Zhenwu M890 supernode instance supports FP8/FP4 low-precision computing. With the ICN Switch 1.0 chip, the Scale-up interconnect scale increases from 16 cards to 64, and card-to-card interconnect bandwidth is raised to 800GB/s. One supernode instance can host inference for trillion-parameter-class MoE large models. Recently, Lingjun supernode instances have opened invitations for testing in the Ulanqab region.
According to media reports, Baidu Kunlunxin has also exhibited a supernode product. Among them, the Kunlunxin 32/64-card supernode is one of the first supernode products in China to achieve mass-production delivery. It uses a self-developed ultra-high-density integration architecture; a single rack integrates 32/64 acceleration cards and can be expanded up to 512 cards.
Bian Ren Technology officially launches the next-generation NPO optical interconnect and distributed decoupling architecture supernode solution, supporting single-supernode 1024-card Scale-up expansion. BR2xx supports FP8/FP4 low-precision, high-compute performance, with higher HBM memory bandwidth and native integration of supernode interconnect capabilities. Its self-developed BLink2.0 supernode interconnect protocol enables up to 1,024 GPUs to share a single memory space.
MudeX (MoX) also officially released its new-generation AI supernode product, Xijing S600. As a system-level compute product built for large-model training, inference, and smart computing center construction, Xijing S600 is comprehensively optimized around high-density deployment, high-speed interconnect, and soft-hard coordination. It can achieve high-speed full interconnect for 64 GPUs within a single server rack, and supports flexible expansion to a ten-thousand-card cluster scale.
In addition, the first fully domestic 100,000-card AI super-heterogeneous cluster—Sugon 8000 (Dengfeng)—full hardware also made its global premiere at WAIC. Previously, Sugon publicly stated that Sugon 8000 (Dengfeng) has been connected to the national supercomputing internet, marking a transition for AIDC from the ten-thousand-card deployment stage to the hundred-thousand-card deployment stage.
China Merchants Securities stated that China-made compute competition is further extending to chip interconnect and system-integration capabilities. As relevant vendors continue to advance product layout for supernodes, China-made supernodes are expected to accelerate large-scale deployment and drive the development of supporting areas such as high-speed connectivity and liquid cooling.
Worth noting is that multiple AI/robot phones were showcased at the event. The phones are expected to upgrade from being just App carriers to running as an agent execution platform, significantly increasing the importance of the device entry point and the runtime base layer. The appearance of Jieyi Agent OS indicates that agents are moving from the application layer further toward the operating-system layer, enabling coordinated operation across mobile phones, industrial equipment, and humanoid robots and other multi-terminal devices.
Meanwhile, embodied intelligence and vertical industry applications are heating up, with industry realization becoming the main storyline. Embodied intelligence scenarios such as humanoid robots, quadruped robots, and AI dexterous hands are concentrated on display. Guotai Haitong Securities noted that the most worth watching at this conference is whether China’s AI industry chain—from underlying computing, chip systems, to large models, Agent OS, intelligent terminals, and on to embodied intelligence and industry applications—can form end-to-end delivery capabilities,推动 the industry from “single-point breakthroughs” to “systematized industrial realization.”
(Science and Innovation Board Daily)