GF Securities released a research report noting that on January 22, the ODCC hosted the 2026 Super Node Conference. The conference focused on a key evolutionary direction for AI computing power: Scale-Up (i.e., super nodes), aiming to address core challenges such as single-machine performance limitations arising from the training and inference of trillion-parameter large models. Focusing on the Scale-Up evolution, the penetration rate of super nodes among cloud providers and computing power providers like Alibaba, Tencent, and Huawei is expected to consistently exceed expectations. Currently, most Scale-Up networks utilize copper cable interconnect solutions, but these face significant limitations in system design and transmission distance. It is recommended to focus on sectors such as switching equipment, NPO, and DAC/AEC. The main views of GF Securities are as follows: The core characteristics of current model inference (long context, high concurrency, real-time interaction) are driving the Scale-Up upgrade. Model inference is gradually evolving to handle more complex multi-step decision-making and engage in multi-turn conversational interactions with users, generating continuous, ultra-long context, real-time inference demands. Long-context inference relies on the KV Cache mechanism, but cached data requires frequent cross-node transmission in distributed environments. If communication bandwidth or latency is insufficient, GPUs may idle while waiting for data, leading to efficiency degradation. Furthermore, these KV caches typically reside in the memory of multiple GPUs, requiring the inference process to continuously access remote memory across cards. In the Scale-Up architecture, the memory of all accelerators appears as a single, shared pool, meeting the high-performance network bandwidth and latency requirements of the inference phase. Domestic super node acceleration is accelerating, focusing on Scale-Up changes. (1) At the Yunqi Conference in September 2025, Alibaba Cloud released its new-generation Panjiu AI Infra2.0 AL128 super node server. The core of Panjiu AL128 addresses GPU interconnectivity, aiming to achieve optimal synergy between computing power and communication in large model training and inference. Compared to traditional architectures, it can boost inference performance by 50% under equivalent computing power. (2) Tencent's ETH-X super node project is a two-step process. It first proposed the "ETH-X" architecture, focusing on optimizing the "execution order" issue in GPU-to-memory communication. Concurrently, it is researching the "ETH-X Ultra" solution, which aims to explore the application of all-optical interconnect for Scale-Up connections. (3) At its Connect conference in September 2025, Huawei announced a 3-year computing power plan. It is developing the Ascend 950 and planning future 960 and 970 series. Among these, a 950-based 8192-card super node is expected to be officially launched in Q4 2026. GF Securities believes that the penetration rate of super nodes among cloud and computing power providers like Alibaba, Tencent, and Huawei is expected to consistently exceed expectations. Investment opportunities in optical interconnect for Scale-Up warrant attention. Most current Scale-Up networks employ copper cable interconnect solutions, but these have significant limitations in system design and transmission distance. Taking NVIDIA's GB200 NVL 72 as an example, its cabinet internally uses copper cables for full connectivity, leveraging the advantages of copper's high reliability, low cost, and low latency. However, limitations exist: (1) From a physical form factor perspective, copper interconnects are constrained by link distance, typically limited to within a single cabinet. This necessitates deploying more chips within a unit space for an AI Rack, and the high-density copper interconnect solution also poses challenges for power supply, cooling, and cabling. (2) As the Scale-Up domain is expected to expand to hundreds or even thousands of cards in the future, copper interconnects become inadequate due to distance limitations. According to data from an Alibaba whitepaper, the transmission distance for active AEC 100G/lane is only 7 meters, whereas optical interconnects can cover tens of meters, hundreds of meters, or even greater distances, highlighting their advantage. Risk warnings include the potential for slower-than-expected AI infrastructure construction; risks associated with slower-than-expected development of AI applications; risks related to changes in import/export policies in the AI field; and risks of reduced international cooperation.
Comments