Sinolink Securities released a research report stating that large model training imposes rigid demands on interconnect bandwidth, making hyper-nodes the core foundation of intelligent computing. With intensive product releases and deliveries from domestic and international manufacturers, the industry is entering a phase of large-scale implementation. The shift to full-rack delivery increases complexity, and leading ODM manufacturers are expected to see a structural improvement in their gross margins. The main views of Sinolink Securities are as follows:

The hyper-node industry is entering a period of large-scale deployment. 1) Large model training and inference have created rigid demand for Scale-up domains. The normalization of trillion-parameter models and MoE architectures, along with the high-frequency All-to-All communication from expert parallelism (EP), places extreme requirements on interconnect bandwidth. Traditional Scale-out clusters face physical constraints in communication, power consumption, and complexity. In this context, Scale-up hyper-nodes, which offer high bandwidth and unified memory addressing, have become the core infrastructure for the new generation of intelligent computing. 2) Supply Side: Domestic and international manufacturers are intensively releasing hyper-node products. The supply side of the industry has entered a period of intensive fulfillment. Overseas, NVIDIA's GB200 NVL72 full racks are operating at full capacity and are accelerating towards the NVL576 configuration. Domestically, the sector is gaining momentum: Huawei's CM384 has entered the engineering delivery phase, while leading manufacturers such as Sugon (scaleX640/40), Baidu (Tianchi 256/512), and Alibaba (Panjiu 128) have successively released or commercialized their rack-level hyper-node products. Strong logic on both the supply and demand sides confirms that the hyper-node industry has moved beyond the technical validation phase and has fully embarked on large-scale implementation.

Cluster-based delivery is driving a structural increase in gross margins for ODM manufacturers. The delivery unit has shifted from individual white-box servers to full racks and even Pod-level systems. Full racks present significant engineering barriers in three dimensions: power supply architecture, cooling, and interconnection, making the R&D and delivery of hyper-nodes a comprehensive challenge that tests a manufacturer's technical, operational, and engineering capabilities. As delivery complexity increases and the number of qualified suppliers remains limited, leading ODMs possess the practical ability to pass on engineering premiums to customers, with gross margins expected to migrate to a higher range.

The design of hyper-node racks must address the challenges of high-density GPU collaboration, with value across multiple segments expected to increase significantly. Core equipment manufacturers, leveraging their strong technological positioning, are entering a cycle of performance realization: 1) Inspur Information has a dual-line layout in the hyper-node matrix, having released the YuanNao SD200 and CRS6000S servers, achieving ultra-high-density deployment of 32/64 domestic AI chips per rack. The 32-card Scale-up high-speed interconnect domain it built increases inter-card communication bandwidth by 8 times, substantially reducing the training cycle for hundred-billion-parameter models from months to weeks. Furthermore, the company is leading the formation of a hyper-node innovation consortium to create the "Beijing Solution," collaborating with the industry chain to tackle interconnect protocols and application implementation, effectively bridging the "last mile" for large-scale hyper-node deployment. 2) Huaqin Technology, leveraging its full-stack design barriers in computing, networking, and liquid cooling, expects its hyper-node project to begin shipments in the second quarter of this year and enter large-scale delivery in the second half. The full-year revenue guidance for 2026 related to this project exceeds 10 billion yuan, accelerating its establishment of a leading industry position.

Risk warnings include intensified industry competition, slower-than-expected progress in technology R&D, and risks associated with cyclical fluctuations in capital expenditure within specific downstream industries.