智通财经网·2026-01-27 07:09Core Insights - The report from GF Securities highlights the significance of the Scale-Up architecture in addressing the challenges posed by training and inference of trillion-parameter models, particularly focusing on the advancements in AI computing power [1][2]. Group 1: Current Trends in AI Computing - The core characteristics of model inference, such as long context, high concurrency, and real-time interaction, are driving the upgrade of Scale-Up systems [2]. - The need for high-performance networks during the inference phase is emphasized, as the architecture allows all accelerators' memory to be presented as a single shared pool, addressing bandwidth and latency requirements [2]. Group 2: Domestic Developments in Scale-Up - Alibaba Cloud introduced the new generation of Panjiu AI Infra2.0 AL128 supernode server, which aims to optimize computing power and communication synergy, achieving a 50% improvement in inference performance compared to traditional architectures [3]. - Tencent's ETH-X supernode project is being developed in two phases, focusing on optimizing GPU and memory communication and exploring full optical interconnection for Scale-Up [3]. - Huawei announced a three-year computing power plan, with the Ascend 950 supernode expected to launch in Q4 2026, indicating a strong commitment to advancing Scale-Up technology [3]. Group 3: Investment Opportunities in Optical Interconnection - Current Scale-Up networks primarily use copper cabling, which has limitations in system design and transmission distance, making it less suitable for future expansions [4]. - The advantages of optical interconnection are highlighted, as it can cover much greater distances compared to copper, which is limited to around 7 meters for certain applications [4].