智通财经APP获悉，中信证券发布研报称，预计超节点方案有望快速上量，作为未来底层AI基础设施的 基本计算单元，超节点scale up域具备高效通信带宽、原生内存语义等优势，天然适配当前主流的MoE 架构模型的计算。超节点在系统层面的"反解耦"使得整体系统价值得以提升，本身设计也面临更多挑 战，多芯片功耗、散热、机柜可靠性等诸多挑战亟待解决，同时也要考虑到系统运行中所面临的困难， 中信证券认为超节点有望通过更高的技术附加带来整机环节价值量的提升。中信证券看好超节点服务器 整机环节的未来发展，建议关注产业链相关公司。 中信证券主要观点如下： MoE架构模型对硬件提出新要求，Scale up超节点应运而生。 在Scaling law的发展背景下，主流AI大模型为追求更大参数规模与更高运行效率，普遍采用MoE（混合 专家）架构。得益于专家网络的独特结构，其天然适配专家并行的计算模式——这种方式虽能有效优化 计算与访存瓶颈，却随之引入了新的通信难题，基于Scale up网络的超节点由此应运而生。相较于传统 八卡服务器，超节点面临更为复杂的系统性挑战：一是海量芯片协同工作带来的系统散热压力；二是多 芯片间光、铜混合互连方 ...

Core Insights - The report from CITIC Securities indicates that supernode solutions are expected to scale up rapidly, serving as fundamental computing units for future AI infrastructure [1] - Supernode scale-up domains possess advantages such as efficient communication bandwidth and native memory semantics, making them well-suited for current mainstream MoE architecture models [1] - The "decoupling" at the system level enhances the overall system value, although the design faces challenges including multi-chip power consumption, heat dissipation, and cabinet reliability [1] - The industry anticipates that supernodes can increase the value of the entire machine segment through higher technological added value, and there is a positive outlook for the future development of supernode server machine segments [1] Industry Implications - The report suggests paying attention to companies within the supply chain related to supernode technology, indicating potential investment opportunities in this sector [1]

港股"A+H"阵营即将迎来一家"超级隐形冠军"。 12月3日，国内数字化解决方案领军企业紫光股份有限公司(以下简称"紫光股份")正式向香港交易所主 板递交上市申请，由中信建投国际、法国巴黎银行及招银国际担任联席保荐人。 该公司是中国领先的ICT基础设施及服务提供商，其发展历程可追溯至1988年，经过三十余载的发展， 公司已成为新一代信息通信领域的领军企业。据弗若斯特沙利文数据，按2024年收入计，紫光股份在中 国数字基础设施市场位列第三，市场份额达8.6%，并在网络基础设施、计算基础设施等细分市场均位 居行业前列。 与此同时，紫光股份在二级市场也"资历不浅"，早在1999年，公司就已在深交所完成上市，截至12月8 日收盘，紫光股份(000938.SZ)股价为25.43亿元，总市值为727亿元。一旦公司成功登陆港交所，其也将 实现"A+H"上市格局。 业绩"AB面"：营收下滑VS利润下滑 据智通财经了解，紫光股份是全球领先的数字化解决方案提供商，该等解决方案包括ICT基础设施产 品、软件及云服务，主要用于人工智能训练及推理以及大数据处理等场景。 据招股书披露，2022年至2024年，紫光股份营收一直保持着稳健增 ...

Market Overview - The A-share market experienced fluctuations, with the Shanghai Composite Index down by 0.97%, the Shenzhen Component Index down by 1.93%, and the ChiNext Index down by 2.82% [1] Sector Performance - Technology stocks saw a general decline, with notable drops in companies like Yizhongtian (Xinyi, Zhongji Xuchuang, Tianfu Communication) and a 4% decrease in SMIC [3] - The gas sector surged, with Shouhua Gas hitting the daily limit up, while the banking sector strengthened, with Industrial and Agricultural Bank of China reaching historical highs [3] - The port and transportation sectors also showed strength [3] Notable Stock Movements - HeFu China, which had a remarkable run of 12 gains in 13 days, attempted to hit the limit up again but closed with a 7.69% increase after a late surge [3] - The company issued a risk warning regarding its stock price significantly deviating from its fundamentals, indicating potential trading risks [3] Renewable Energy Sector - The photovoltaic sector saw activity, with Longi Green Energy experiencing a peak increase of over 8%, closing at 21.88 CNY per share, with a market capitalization of 165.81 billion CNY [5] - Longi Green Energy announced plans to acquire approximately 61.9998% of Suzhou Jingkong Energy Technology Co., indicating a strategic move into the energy storage sector [7] Energy Storage Market Insights - According to Guoxin Securities, the domestic energy storage system bidding has seen a significant increase, with a cumulative bidding scale of 166.3 GWh from January to October, representing a year-on-year growth of approximately 172% [7] - New energy storage installations reached 85.5 GWh in the same period, a 71% increase year-on-year, with a notable 224% increase in new project filings in October [7] GPU and AI Sector Developments - The GPU sector saw a boost with the announcement of the IPO intentions of Moer Thread, aiming to become the first domestic GPU stock, leading to a rise in related stocks like Lihexing and Heertai, which increased by over 5% [8] - ZTE Corporation introduced a super node server supporting 64 GPUs, enhancing its capabilities in large-scale intelligent computing clusters [11] - Minsheng Securities highlighted the potential of super nodes in cluster expansion and the focus on aerospace computing as a key area for domestic computing development [11]

Core Insights - The report emphasizes a positive outlook on the AI computing sector, particularly for domestic computing supernodes, despite ongoing trade tensions between China and the US [5][6] - Significant capital expenditure is anticipated in the AI sector, with OpenAI's partnerships with NVIDIA and AMD indicating a substantial increase in computing power and investment [5] - The domestic supernode server market is expected to see significant growth by 2026, driven by advancements from major companies like Huawei and Alibaba [5][6] Market Trends - The report notes fluctuations in major domestic market indices, with the Shanghai Composite Index closing at 3,912.21, up 1.22% [4] - The communication sector is highlighted as a key area of focus, with ongoing developments in overseas optical modules and domestic computing supernodes [5] Industry Commentary - The report discusses the impact of recent trade negotiations, suggesting that market participants should remain calm and consider increasing positions during market dips [5] - It highlights the strategic partnerships formed by OpenAI with major tech companies, which are expected to drive demand for optical modules and related technologies [5][6] - The report identifies key players in the optical module and supernode server markets, recommending companies such as ZTE, Inspur, and Unisoc for potential investment opportunities [7][8]

Group 1 - North American CSP cloud providers, including Amazon, Google, Meta, Microsoft, and Oracle, plan to invest over $370 billion in capital expenditures for the fiscal year 2025, representing a 40% year-on-year increase [1][3] - Oracle's latest quarterly report shows that its remaining performance obligations (RPO) reached $455 billion at the end of FY26Q1, an increase of $317 billion from the previous quarter, with expected capital expenditures of $35 billion for FY2026 [1][3] - The global Ethernet switch market is projected to generate $14.5 billion in revenue in Q2 2025, reflecting a year-on-year growth of 42.1% [1][3] Group 2 - In September, the communication sector underperformed the broader market, with the communication index declining by 0.15% while the CSI 300 index rose by 3.20%, ranking 14th among 31 primary industries [2] - The PE valuation for major companies in the communication sector was 23.8 times in September, slightly below the historical median of 33.6 times over the past decade [2] - Notable stock performances in September included Tengjing Technology (+66%), Chunzhong Technology (+92%), and Yuanjie Technology (+44%) [2] Group 3 - Domestic CSP cloud providers in China, such as ByteDance, Alibaba, Tencent, and Baidu, are expected to invest over 450 billion yuan in AI computing power by 2025, with accelerated industrialization of domestic super-node data centers [4] - Major chip manufacturers like NVIDIA, AMD, and Broadcom are increasing collaboration with CSP cloud providers, leading to a rise in demand for high-end chips [4] - Investment recommendations include focusing on AI computing infrastructure development, particularly in optical devices, communication equipment, and liquid cooling technologies [4][6] Group 4 - The three major telecom operators in China are considered important assets for dividend allocation, with stable operations and increasing dividend payout ratios, suggesting a long-term investment strategy [5][6] - The recommended stock portfolio for September 2025 includes China Mobile, Zhongji Xuchuang, ZTE, and Guanghetong [6]

Core Insights - The core viewpoint emphasizes that AI will replace energy as the most important commodity, driving daily operations across various industries, with computational power being the key element for AI's industrial application [1][12] - The "computational power question" is central to the AI industry, focusing on how to make computational power more accessible and transition from a "technology race" to "industry application" [2] Group 1: Full-Stack Intelligent Computing - ZTE Communications proposed a full-stack open intelligent computing solution that covers hardware systems and AI platforms, ensuring compatibility with mainstream GPU and CPU, allowing businesses to flexibly choose based on their needs [3] - The solution includes self-developed switching chips, intelligent computing servers, super node servers, and AiCube integrated computing machines, enabling efficient training and inference [3] Group 2: Super Node Servers - The super node servers showcased can support 64 GPUs per machine, allowing for vertical and horizontal scaling to meet extreme performance needs, thus supporting millions of user requests and complex predictive models [5] - This solution enables businesses to gradually upgrade their computational power as their operations evolve, rather than requiring an all-in commitment from the start [5] Group 3: Intelligent Computing Centers - The intelligent computing center focuses on energy efficiency, with a modular and prefabricated architecture that reduces delivery cycles by 40% and significantly lowers energy consumption [7] - As of June 2023, China's operational computing center rack scale reached 10.85 million standard racks, with intelligent computing capacity growing by approximately 40% compared to 2023 [6][7] Group 4: Collaborative Ecosystem - The partnership between ZTE Communications and Alibaba Cloud has evolved over 13 years, expanding from the Chinese market to a global presence, indicating the importance of collaboration in advancing AI industry [8] - The collaboration aims to create a symbiotic ecosystem that integrates software and hardware, supply and demand, and domestic and international markets [8][9] Group 5: Global Solutions - ZTE Communications and Alibaba Cloud have established cloud capability centers in various countries, promoting a "joint development + local delivery" model that meets local regulations and market needs [11] - This approach transforms "Chinese-style computing engineering" into globally applicable solutions, supporting enterprises in their international operations [11][12]

Core Insights - The global cloud service providers (CSPs) are accelerating their AI arms race, benefiting the computing infrastructure from AI development, which has become a main track in the telecommunications industry [1] Group 1: AI Infrastructure Developments - Alibaba Cloud has launched the new generation of the Panjiu 128 ultra-node AI server, featuring an orthogonal architecture design that supports 128 AI computing chips per cabinet, with a power density of 350 kW and single chip high-load operation of 2000 W, achieving a hardware fault prediction accuracy of 99% [1] - ZTE's ultra-node server supports 64 GPUs and possesses both Scale-Up and Scale-Out expansion capabilities [1] Group 2: Optical Communication Innovations - Alibaba's self-developed 102.4T switch supports 800G/1.6T optical modules, and the UPN network system utilizes LPO/NPO optical interconnect technology, expected to reduce costs by over 30% [1] Group 3: Investment Opportunities in AI - The Guotai AI ETF (159388) tracks the ChiNext AI Index (970070), which has a fluctuation of 20%, selecting listed companies involved in software development, smart hardware manufacturing, and big data processing related to AI technologies and applications [1] - This index focuses on the innovation capability and growth potential of China's emerging technology industry, encompassing representative enterprises in the AI field [1]

Core Insights - The AICC2025 Artificial Intelligence Computing Conference was held in Haidian District, focusing on AI infrastructure and optimization of domestic AI computing systems [1] - The conference introduced the "Beijing Plan" for super-node intelligent computing applications, emphasizing the need for advanced computing resources in the era of large models [1] Group 1: AI Infrastructure and Development - The demand for computing power has surged due to the explosive growth in training and inference of large models, making traditional computing architectures inadequate [1] - Over 30 companies and institutions, including China Mobile and Inspur Information, collaborated to release the "Beijing Plan," which aims to integrate AI chips, system manufacturers, and application innovators [1] Group 2: Super-node Servers and Industry Applications - Super-node servers consist of dozens to hundreds of AI chips, functioning as a compact computing unit with low-latency interconnect technology [2] - The goal is to make AI technology accessible across various industries, moving from exclusive use by leading players to widespread application [2] Group 3: Policy Support and Innovation Centers - A new innovation center was established in Beijing to focus on reconfigurable computing chip architecture and software ecosystem optimization, involving companies like Qingwei Intelligent and Tsinghua University [2] - The conference featured a policy consultation desk, offering guidance on a subsidy program for AI companies, with a maximum subsidy of 20 million yuan to support breakthroughs in models and intelligent agents [2]

Group 1 - The core of the "Beijing Plan" is to develop a super-node innovation consortium that integrates companies from systems, large models, and applications to create industry-specific "intelligent bodies" [3] - The intelligent body proliferation is a key goal of the national "Artificial Intelligence +" initiative, aiming for a penetration rate exceeding 70% by 2027 and 90% by 2030 [3] - The "Beijing Plan" will leverage domestic super-node platforms to develop solutions for various scenarios such as scientific research, embodied intelligence, healthcare, intelligent manufacturing, and education [3] Group 2 - In 2024, Beijing's AI core output value is projected to reach 350 billion yuan, with 159 registered large models accounting for 30% of the national total [3] - The city has gathered over 2,400 companies and a large number of top talents, with unicorn companies making up half of the national total [3]