Dongxing: The rapid development of super nodes and Scale-up networks. The market will reevaluate the value of domestic super node sectors.

Dongxing released a research report stating that super nodes and Scale-up networks are key infrastructure to break through the computational power and communication bottlenecks, supporting trillion-level large models and high real-time applications. Top AI chip manufacturers such as NVIDIA, Google, AMD, and Huawei are laying out in this field. The firm believes that super nodes and Scale-up networks are in a rapid development phase and will become important application markets for emerging technologies such as computational chips, network components (PCB boards, switching chips, optical devices, high-speed copper cables), storage components, power supply, and cooling facilities. In addition, the firm believes that in the development of super node technology, the market will continue to reassess the value of Google, AMD, and domestic super node sectors. Dongxing's main points are as follows: (1) NVIDIA: Leading advantage of super nodes established on NVLink and NVLink Switch In terms of super node technology solutions, NVIDIA has a leading advantage. Between 2024-2025, NVIDIA will gradually launch mature super node solutions such as GH200 NVL72, GB200/ GB300NVL72. According to Morgan Stanley's forecast, in 2025, NVIDIA's GB200/300 NVL72 shipments are estimated to be about 2800 units. Looking ahead to 2026-2027, NVIDIA plans to release Vera Rubin NVL144 and Rubin Ultra NVL576. The interconnection of GPU numbers will further develop from 72 to 576. At that time, NVIDIA will introduce the new generation Kyber rack, introducing the NVLink Switch Blade architecture (NVLink switch blade) to replace traditional 5000+ active copper cables with PCB midboards. It can be seen that Rubin Ultra NVL576 still maintains strong engineering innovation capabilities. NVIDIA's advantage in super nodes is built on NVLink and NVLink Switch. To achieve high bandwidth and low latency data transmission for AI training clusters, NVLink has redesigned the communication architecture and introduced a series of advanced technologies, including mesh topology, differential signal transmission, flow scheduling credit mechanism, multi-lane binding technology, unified memory space, etc. By 2025, NVLink 5 Switch supports a bandwidth of 1800GB/s per GPU to GPU, enabling a domain of 72 GPUs with a total bandwidth of 130 TB/s (bidirectional) and supporting full interconnect communication among 72 GPUs. In subsequent plans, the GPU-to-GPU communication bandwidth of NVSwitch Gen6 and Gen7 continues to increase to 3.6TB/s. However, on the other hand, the rise of Scale up networks originated from meeting the needs of tensor parallelism (TP) and expert parallelism (EP) in large-scale model distributed training and inference. Currently, the AI industry is also exploring technical solutions to reduce the scale of TP and EP, thereby lowering the upper limit of Scale up network scale. The firm believes that the development space of Scale up networks may limit NVIDIA's leading advantage in the super node field. In order to maintain its leading position, the integration of Scale up and Scale out networks may become a new development trend for NVIDIA super nodes. (2) Huawei: Openly promoting the Liqiu Interconnect Protocol, catching up with NVIDIA in super node performance The domestic Scale Up protocol has not been unified, and Huawei's Liqiu protocol has not been widely accepted in the domestic industry. In terms of the Scale Up protocol, Huawei has introduced the Liqiu protocol and shifted towards open standards starting from version 2.0. In addition, other domestic manufacturers are exploring various interconnect protocols, including China Mobile's OISA, Tencent's ETH-X, Qualcomm's ETH+, and ZTE Corporation's OLink. In order to break down ecosystem barriers, efforts are being made domestically to promote standardization, such as the Ministry of Industry and Information Technology leading the promotion of the CLink protocol to establish a unified domestic standard. Huawei's super nodes rely on a clustered approach to achieve performance parity. The Atlas 950 super node is expected to be released in the fourth quarter of 2026, with a computing power of 8 EFLOPS (FP8), compared to NVIDIA's NVL144 which is also planned to be released in the second half of 2026 with a total computing power of 2.52 EFLOPS (FP8). Furthermore, the Atlas 950 super node has a significantly larger memory capacity of 1152TB and an interconnect bandwidth of 16.3PB/s, leading in these aspects. The firm believes that in the short term, Huawei's super nodes will achieve performance parity through clustering, but balancing the complexity, reliability, and power consumption dimensions of super nodes is necessary. In terms of overall solutions, NVIDIA's advantages in super node chip technology, software ecosystem, and system integration remain difficult to shake. The interconnect solution of the Atlas 950 super node may be adjusted, showing that Huawei's super node technology still needs to be solidified in the standardization phase. Compared to the previous generation super nodes, the Huawei Atlas 950 super node no longer uses a fully optical interconnection architecture; instead, it adopts a hybrid design of "cabinet internal orthogonal copper interconnect + cabinet external optical interconnect" to achieve high reliability, low cost, and low power consumption connections within the cabinet, while ensuring system scalability through optical interconnection between cabinets, effective control of the total cost of ownership (TCO). (3) Google: Establishing optical interconnect super nodes, forming asymmetric competition with NVIDIA Google's TPU super nodes have established a mature optical interconnect Scale up network. In terms of technological maturity, Google will gradually launch three generations of TPU super nodes, TPU v4, TPU v5p, and TPU v7, between 2023-2025, completing technology exploration and standardization. Furthermore, TPU v7 has also received recognition from external enterprises. In 2026, Anthropic will directly purchase nearly 1 million TPU v7 Ironwood AI chips from Broadcom and deploy them locally in its controlled data centers. In 2027, Google will release the 8th generation TPU, benchmarking against Nvidia's Vera Rubin. It can be seen that by then, the performance metrics of Google's TPU super nodes will be further optimized and improved. Google's TPU super node competitive advantage is built on OCS switches, with a unique technological route. Compared to other super node manufacturers such as NVIDIA, Huawei, and AMD, Google is the first enterprise globally to deploy optical circuit switches (OCS) on a large scale in Scale up networks, with a unique technological route. Google's OCS switches involve deep crossover applications of precision optics, mechanical engineering, and semiconductor processes, building a high barrier of technological moat in the optoelectronic interconnect field. Optical circuit switching technology has many advantages compared to electronic packet switching: optical circuit switches can reuse multiple generations of optical transceiver module technology, the energy consumption per bit of optical circuit switches is several orders of magnitude lower than that of electronic packet switches, and optical circuit switches introduce minimal delays. Commercial deployment of OCS switches faces multiple challenges: optical circuit switches need to expand to support hundreds of ports to accommodate a sufficient number of NPU interconnections; constrained by the control software and reflection mirror configuration delays of optical circuit switches, the switching delay of commercial optical circuit switches is usually 10-20 milliseconds; to reduce link power, the insertion loss of optical circuit switches needs to be controlled at an ideal level. In order to build a cost-effective and large-scale optical switching layer, Google has innovatively developed three core hardware components: optical circuit switches, wavelength division multiplexing optical transceiver modules, and optical ring transformers. The optical core module of Google's Palomar optical circuit switch is MEMS micro mirrors that achieve optical routing functionality; the wavelength division multiplexing optical transceiver module improves wiring efficiency and supports the key to large-scale and continuous expansion of data centers; the optical ring transformer is the core device to realize bidirectional communication of optical circuit switch links, reducing the number of optical circuit switch ports and fiber by half. (4) AMD: UALink becomes an important open standard, super nodes have the potential to become a strong competitor to NVIDIA As an open technology route for Scale up networks, UALink has become a significant standard. The 1.0 version of the specification was officially released in 2025; it is expected that the UALink 2.0 version will be released in 2026. The firm believes that UALink is currently transitioning from the standard-setting stage to the product implementation stage, and it is expected that the UALink ecosystem will experience a breakthrough development by 2027, being accepted by numerous data centers. Currently, the UALink alliance has broad industry support, with over 100 member units as of the end of January 2026, and it is expected to challenge NVIDIA's NVLink advantage. AMD's Helios frame super node is expected to become the mainstream choice in the industry. The Helios frame adopts a dual-width frame design, increasing the width from one rack to two racks, achieving a good balance between complexity, reliability, and performance. In terms of computing power, memory, interconnect bandwidth, etc., the MI455x series Helios cabinet is currently the most competitive product against NVIDIA's NVL72 cabinet; and in the power consumption sector, the Helios frame has significant advantages compared to the GB200 NVL72 cabinet. Additionally, the dual-width structure reserves physical space for future upgrades, such as expanding to a 144-GPU configuration without the need to redesign the rack infrastructure. Investment Strategy Starting from 2025, super nodes have become an important direction for technical innovation in AI computing networks. From the perspective of AI infrastructure competition, AI chip manufacturers have shifted their competition from chip computing performance to a dual battlefield of chip + Scale up network. Therefore, in addition to traditional chip companies like NVIDIA, Huawei, AMD, and Google, more global companies are joining the competition in the super node race, including Microsoft, Meta, Amazon, China Mobile Limited, Alibaba, ByteDance, Tencent, Baidu, Dawning Information Industry, ZTE Corporation, Inspur Electronic Information Industry, Unisplendour Corporation (Unisplendour), Hygon Information Technology, Muxi Share, EmbedWay Technologies, and others. The global super node competitive landscape has not been established yet. NVIDIA currently holds a leading position, but the continued efforts of giants like Google, AMD, and Huawei in the super node field have posed a challenge to NVIDIA's dominant position. From the stock price performance perspective, in 2023-2024, NVIDIA's stock price significantly outperformed Google, AMD, and the A-share China Computing Power Index. However, in 2025, NVIDIA's stock price had a cumulative increase of 38%, significantly lagging behind Google, AMD, and the A-share China Computing Power Index. The firm believes that in the development of super node technology, the market will continue to reassess the value of Google, AMD, and the domestic super node sector. Investment Recommendations (1) Positive outlook for Google, AMD, and domestic super node manufacturers; (2) Positive outlook for the supply chain of NVIDIA, Google, and AMD super nodes, including PCB backplanes, high-speed copper cables, optical modules, power supply, and liquid cooling systems; (3) Based on switches and chips being key devices for Scale up network interconnection, positive outlook for suppliers of core components of Google's optical switch (OCS) and switch chip developers under the UALink standard. Risk Warning: (1) Changes in LLM training and reasoning technology paths; (2) Super node performance and power consumption require balancing; (3) Lower-than-expected shipment volumes of super nodes from various manufacturers due to supply chain impacts; (4) Slower-than-expected growth in AI application endpoints.