Dongwu Securities: Since 26Q2, supernodes have intensively released domestic computing power and entered the system-level verification stage

The Zhitong Finance App learned that Dongwu Securities released a research report saying that domestic supernodes are expected to enter the first year of commercialization in 2026. They have initially formed a complete chain of product supply, public procurement, large-scale deployment, and cloud service output. The industry's focus is on whether to deliver stably, operate continuously, and reduce unit token costs. The whole machine platform was the first to benefit, and core components and infrastructure opened up incremental space. The proposal focuses on complete platforms with supernode product development, system integration and large-scale delivery capabilities, as well as domestic computing power chips, high-speed switching chips, high-speed connectors, liquid cooling and cooling, and power supply support.

The main views of Dongwu Securities are as follows:

Domestic supernodes are expected to enter the first year of commercialization in 2026

Since 26Q2, domestic supernodes have been released intensively, and Lenovo, Inspur, Huawei, ZTE, Mu Xi, Suiyuan, and Alibaba Cloud have successively launched supernode solutions or cloud services; at the same time, Zhongke Shuguang built China's first nationally produced 100,000-card AI supergroup, and domestic computing power infrastructure expanded rapidly from single-cabinet products to hyperscale clusters. The demand side simultaneously initiated scale verification. China Mobile collected 6,208 AI acceleration cards and 776 computing nodes, and Huawei revealed that the total deployment of Shengteng 384 supernodes exceeded 750 sets. Unlike previous technology releases and prototype displays, in 2026, domestic supernodes have initially formed a complete chain of product supply, public procurement, large-scale deployment, and cloud service output. The industry's focus is shifting to whether it can be delivered stably, continuously operated, and reduced unit token costs.

What are the differences between Chinese and foreign supernode solutions?

In overseas markets, Nvidia relies on the CUDA ecosystem, NVLink and Blackwell platforms to build a highly vertically integrated system. The GB300 NVL72 integrates 72 BlackWell Ultra GPUs in a single cabinet, and the total NVLink bandwidth reaches 130TB/s; AMD Helios chose an open route, connecting 72 Mi455xGPUs through OCP, UALink, and UEC standards, emphasizing ecological compatibility and open supply chain. Domestic manufacturers mainly compete around autonomous ecosystems and system integration capabilities. Huawei and Alibaba Cloud rely on self-developed AI chips, interconnection architectures, basic software, and cloud services to form an end-to-end solution. Among them, the Huawei CloudMatrix 384 supernode has pooled 384 NPU resources; OEM manufacturers such as ZTE, Lenovo, Inspur, and Xinhua focus more on adapting multiple domestic GPUs, delivering machines, and implementing industry scenarios.

Why will supernodes be launched at an accelerated pace in 2026?

Model load upgrades and domestic computing power requirements are jointly driving the accelerated implementation of supernodes. On the training side, the MoE model broke through scale bottlenecks through expert parallelism, but all-to-all communication significantly increased the pressure on data exchange between GPUs; the inference side, long context, and high concurrency continued to increase KVCache capacity, and prefill/decode separation further increased KV Cache transmission requirements across nodes. Traditional 8-card servers mainly rely on scale-out network expansion. As the cluster scale expands, cross-node communication delays and network congestion gradually become bottlenecks. The supernode integrates more acceleration cards into a unified computing domain with high bandwidth and low latency interconnection, reducing data handling and communication waiting across nodes, and improving GPU utilization.

How do major internet companies arrange supernodes?

Major Internet companies have not only large-scale training and inference workloads, but also self-developed chips, cloud platforms, and model service capabilities. They can use internal business to complete software and hardware optimization and product verification, and then export computing power to external customers. Alibaba Cloud launched the Panjiu AL128, which integrates 128 Zhenwu M890 and ICNSwitch in a single cabinet; Baidu launched the Tianchi Super Node, which uses 32 Kunlun XPU to achieve a full layer of interconnection, and combines the Baige platform to optimize communication, operators, and inference scheduling; Volcano Engine launched a 102.4T self-developed switch and HPN 6.0 network for 100,000 card GPU clusters. Major Internet companies are deploying supernodes not only to expand the supply of computing power, but more importantly, to improve the utilization rate of self-developed chips and clusters, reduce model training and inference costs, and transform internal infrastructure capabilities into computing power leasing, model calling, and enterprise AI service revenue.

Risk warning: Industry competition intensified, CSP capital expenditure and AI computing power requirements fell short of expectations, and customer verification and scale delivery progress fell short of expectations.