AI computing power expansion path analysis reshapes the industrial chain of super nodes! Shenwan Hongyuan Group: Pay attention to Hygon Information Technology, Lenovo and others.

Recently, Shenwan Hongyuan Group Huang Zhonghuang team released a research report, stating that in the current situation where the parameters of large models are growing exponentially, the demand for computing power is accelerating from single point to system-level integration. Under this trend, Scale-up and Scale-out have become the two core dimensions of computing power expansion. Ultra-nodes are not only making progress in cabinet-level interconnection and cross-cabinet networking technologies, but also reshaping the division of labor in the computing power industry, giving rise to investment opportunities in server integration, incremental optical communication, and increased liquid cooling penetration. Using the metaphor of cargo ships, when the total capacity demand expands, Scale-up is to build larger cargo ships, while Scale-out is to increase the number of cargo ships. Scale-up pursues tight hardware coupling; Scale-out pursues elastic expansion to support loose tasks (such as data parallelism). There are essential differences between the two in protocol stack, hardware, and fault tolerance mechanisms, resulting in different communication efficiencies. Currently, Scale-up has broken through the traditional limitations of a single server or cabinet and entered the era of "super nodes." Scale-up can be understood as increasing the number of GPUs within a single node (from 2 cards to 8 cards); but its core lies in achieving full interconnection of GPUs within the node, rather than physically existing in a single server or cabinet. With the evolution of interconnection technology, Scale-up is breaking through the limitations of a single server or cabinet, allowing "super nodes" to span servers and cabinets. The super node is actually the scale-up of the computing network system at the level of a single or multiple cabinets, with mainstream communication solutions within nodes being copper connections and electrical signals, while optical communication is considered for cross-cabinet connections; its hardware boundary with Scale-out is the NIC network card, with external devices such as optical modules and Ethernet switches. The two have significant differences in architecture design, hardware devices, and protocol standards. Currently, Scale-up and Scale-out have not yet fused or crossed over, with chip manufacturers such as NVIDIA, Broadcom, Huawei, and H3C focusing on deepening their efforts in Scale-up, while Ethernet (e.g., Broadcom network chips, Hisilicon network chips, Shengke communication, etc.) are focusing on Scale-out. Under the trend of cabinets/super nodes, AI chip manufacturers vertically integrate to enhance their own communication, storage, and software capabilities, which is a definite trend. Chip giants are strengthening their layout in the computing power network. NVIDIA, AMD, and domestic Hygon Information Technology absorbing and merging Dawning Information Industry are all taking action. NVIDIA has completed 8 mergers and acquisitions in the past 6 years, focusing on the integration of the entire computing power chain. By acquiring network technology (Mellanox), software-defined networking (Cumulus), industry applications (Parabricks), cloud services (Lepton AI), and AI development tools (Run.ai, Deci), they are building a closed-loop ecosystem from chips to applications to compete with cloud giants and penetrate new markets. Hygon Information Technology released a pre-announcement, intending to absorb and merge Dawning Information Industry, which also confirms this industry trend. From an industry synergy perspective, Hygon Information Technology's main source of revenue is CPU+DCU, while Dawning Information Industry's main source of revenue is servers+cloud infrastructure. After the merger, Haiguang+Shuguang will complete the hardware full industry chain deployment from chips to cloud, with significant synergies. So, are there any risks for server manufacturers? Firstly, AI chip manufacturers will not enter the foundry business. After acquiring ZT System, AMD divested its foundry business to avoid competition with OEM/ODM companies, and Haiguang's acquisition of Shuguang is also aimed at strengthening collaboration and enhancing liquid cooling and software capabilities. However, the industrial division of labor in the computing power chain may further refine. Under the trend of super nodes, the interconnection between AI chips and between AI chips and switch chips often requires the use of boards (especially for electrical signal interconnection). Taking NVIDIA as an example, they design their boards internally when products are initially launched, but once the products stabilize, they will open up to OEM partners. At this point, the ability to design boards becomes a core differentiating capability for acquiring more value. Therefore, the division of labor in the foundry industry chain may further differentiate into board design foundry suppliers and cabinet foundry suppliers. In terms of industry opportunities, focusing on the industry chain opportunities under the evolution of technology paths, focusing on hardware interconnection and scenario adaptation, it is recommended to pay attention to AI chip and server suppliers: Hygon Information Technology, Dawning Information Industry, Inspur Electronic Information Industry, Unisplendour Corporation, Digital China Group, LENOVO GROUP, Huaqin Technology, etc.