Orient: 2026 is the year of mass production of domestic super nodes, with continued growth in power consumption leading to increased demand for PSU, HVDC, and SST.

Orient released a research report stating that the super node is the "spear" of domestic computing power attack, and will make up for the insufficient performance of single cards. 2026 will be the year of mass production of domestic super nodes. Inside the super node cabinet, a new Scale up domain brings the demand for exchange chips and exchange nodes; high-speed interconnection is required between exchange nodes and computing nodes, leading to the demand for copper cables, PCB backplanes, and optical interconnects; the power consumption of the super node cabinet has greatly increased, making liquid cooling a necessity, gradually entering the era of full liquid cooling; the super node cabinet involves multiple links, raising the industry entry threshold, increasing the number of value retention links, and prompting a reevaluation for server ODM manufacturers; the power consumption of the super node continues to increase, restructuring the power supply architecture, leading to demand for PSU, HVDC, and SST. Key points from Orient: A must for the development of AI: AIInfra is heading towards the era of super nodes Some investors believe that the super node is just a simple stack-up from eight cards to more computing cards, while Orient believes that the super node is a must for the development of computing infrastructure on both the training and inference sides, and is the cornerstone of future clusters with millions of cards. (1) Training side: With the increase in model parameters and the transition to MoE architecture, TP and EP bring high bandwidth and multi-frequency interconnection demands, and super node servers can effectively support parallel computing tasks by internal high-speed bus interconnection, shortening the training cycles of large models. (2) Inference side: In the age of raising lobsters, the inference tasks require high KVCache cache, and the super node effectively solves the memory wall issue, and generates higher Tokens per single W per second, making it more cost-effective on the inference side. (3) A cluster of millions of cards, Scale up first, super nodes are the foundation of future large-scale clusters. Scale up protocols are moving towards open source and openness, and super nodes help domestic cards to "overtake on the curve" Super nodes can be divided into whole-cabinet super nodes, sub-cabinet super nodes, and Matrix super nodes (cascade super nodes), mainly composed of computing nodes, exchange nodes, TOR switches, power supply units, power supply busbars, cable trays/orthogonal backplanes, liquid cooling heat dissipation kits, etc. Manufacturers represented by Broadcom and Alibaba are gradually launching open standard protocols, breaking the relatively closed competition pattern of Scale up protocols and accelerating the construction of the super node ecosystem. Although domestic chips are slightly behind in process technology, taking Huawei's 384 and GB200NVL72 as examples, the performance of a single Ascend 910C chip in BF16 mode is only one-third of that of the GB200 module. However, through the super node method, the overall performance of a single 384-cluster BF16 is 1.7 times that of the NVL72. Five major trends brought by super node servers (1) The addition of Scale up demand for super nodes, as the Scale up domain continues to expand, driving the continuous improvement of exchange chip ratios; (2) high-speed interconnection is required between exchange nodes and computing nodes, and with the continuous increase in computing card numbers, gradually driving the demand for PCB backplanes and orthogonal connectors; (3) the high-density development of super node cabinets, with continuous power consumption driving the demand for 100% full liquid cooling; (4) from 8-card machines to super nodes, involving networks, power supply, cooling, and other multiple links, raising the industry entry threshold, increasing the number of value retention links, and prompting a reevaluation of value for server ODM manufacturers; (5) super nodes reconstruct the power supply architecture, leading to demand for PSU, HVDC, and SST. Risks and warnings AI development falls short of expectations, industry competition intensifies, geopolitical risks.