联合研究丨行业深度 [Table_Title] 云计算 50ETF 新华联接： 聚焦 AI 技术周期下半场的核心环节 %% %% %% %% research.95579.com 1 丨证券研究报告丨 报告要点 [Table_Summary] AI 技术创新周期可分为上下半场：AI 上半场主要聚焦于模型和方法的创新，通过开发新的算法 和模型架构追求模型的智力极限，算力为王；下半场转向定义问题，基于现有模型能力通过系 统化工程手段推动 AI 与现实场景融合落地变现，应用为王。中证云计算 50 指数全面覆盖云计 算全产业链，软硬件均衡配置，以"算力底座+云端应用"双轮驱动，同步把握 AI 算力基建红 利与软件端成长机会。云计算 50ETF 新华（560660）紧密跟踪标的中证云计算 50 指数；新华 中证云计算 50ETF 联接（A 份额：025790.OF，C 份额：025791.OF）基金主要通过投资于目 标 ETF 来实现对标的指数的紧密跟踪，追求与业绩比较基准相似的回报。 分析师及联系人 [Table_Author] 宗建树 卢之晗 刘胜利 SAC：S0490520030004 SAC：S04905251 ...

Core Insights - The AI era is facing a critical challenge with the shortage of high-speed interconnect networks, which is essential for the efficient operation of large-scale computing clusters [1][10] - Domestic companies are making strides in developing their own computing chips, but the core technology for high-speed interconnects remains dominated by Nvidia, posing a significant risk to the industry [1][10] Group 1: Industry Challenges - The transition from GPU to high-speed interconnects is becoming a new bottleneck as the scale of computing clusters increases from thousands to tens of thousands of nodes [1][4] - The communication time in distributed training can account for 30-50% of the total time, indicating that a significant portion of the investment in computing power is wasted on data transfer rather than computation [4][5] - The demand for high-speed networks has increased by 10 to 20 times as servers now require multiple network cards to support GPU-centric architectures [6] Group 2: Technological Landscape - There are two main technological routes in the high-speed network domain: RoCE and InfiniBand, with the latter being the preferred choice for high-performance computing due to its superior performance metrics [7][10] - InfiniBand networks are used in approximately 60% of the world's high-performance computing systems, and they are almost standard in the largest AI training clusters [10] Group 3: Domestic Developments - In response to the challenges posed by the dominance of foreign technology, companies like Zhongke Shuguang have developed their own high-speed network solutions, such as the scaleFabric, which is fully self-developed [2][11] - The decision to pursue a fully self-developed InfiniBand system was driven by the inadequacy of available commercial IPs and open-source solutions to meet the performance and reliability requirements for large-scale clusters [12]

Investment Rating - The report maintains a "Buy" rating for the computer industry [4] Core Insights - The "14th Five-Year Plan" emphasizes the core strategic position of artificial intelligence (AI) in national development, with the term "artificial intelligence" appearing 30 times, compared to only 6 times in the previous plan [6][8] - The next five years will see AI Agents as the driving force for economic transformation, with a focus on high-value AI Agent growth leading to significant value creation [6][10] - The demand for intelligent computing power is expected to rise significantly, with projections indicating that by 2028, intelligent computing power will account for over 95% of total computing power in China [6][12] - The report highlights the emergence of "Token inflation" due to the rapid growth in AI model usage, with a projected annual Token consumption increase from 0.0005 PetaTokens in 2025 to 152,667 PetaTokens by 2030, reflecting a CAGR of 3418% [6][24] - Investment opportunities are identified in AI-native application companies, edge AI technologies, domestic computing power chain replacements, and collaborative infrastructure for computing and electricity [6][38] Summary by Sections Section 1: The "14th Five-Year Plan" as a Key Period for Intelligent Economy - The plan introduces the concept of "intelligent native," suggesting AI may become a new production factor [11] - The intelligent economy will drive the reconstruction of AI factor value [13] Section 2: Outlook for the "14th Five-Year Plan" - The intelligent economy is set to initiate a rapid explosion in Token usage, with AI Agents transitioning from cost centers to profit centers [17][38] - The report anticipates a significant increase in the number of active AI Agents, from approximately 28.6 million in 2025 to 2.216 billion by 2030, with a compound annual growth rate (CAGR) of 139% [24] Section 3: Comprehensive Upgrade of AI Factors During the "14th Five-Year Plan" - The report emphasizes the importance of high-quality data sets as a core barrier for building irreplaceable AI Agents [16] - The demand for high-quality, proprietary data sets is expected to surge, with a focus on transforming data resources into valuable assets [16] Section 4: Investment Recommendations - The report suggests focusing on AI-native application companies capable of generating scalable revenue, as well as companies that integrate AI Agents with vertical industry know-how [6][38] - Specific companies to watch include Horizon Robotics, JingTai Holdings, Meitu, and others [6]

行业点评 | 计算机 国产 RDMA 技术实现突破，助力超节点加速落地 RDMA（Remote Direct Memory Access，远程直接内存访问）作为一种网 络通信技术，致力于解决大规模并行计算中的数据传输延迟和 CPU 消耗问 题。在生成式 AI 的时代，RDMA 已成为 AI 算力基础设施重要的底层技术。 中科曙光于 2026 年 3 月 12 日正式发布首款全栈自研 400G 无损高速网络— —scaleFabric，实现了国产 RDMA 技术重要突破。scaleFabric 作为国内首 款原生无损 RDMA 高速网络，基于原生 RDMA 架构，在 112G SerDes IP、 交换芯片、网卡到交换机、驱动与管理软件等关键技术上实现自主研发。 scaleFabric 面向超大规模智算集群设计，有望为超大规模智算集群提供高带 宽、低时延、真无损、超可靠的底层基础。 建议关注：1）算术集成：中科曙光（已覆盖）。2）AI 芯片：寒武纪（已覆 盖）、海光信息（已覆盖）、天数智芯。3）互联技术：盛科通信（已覆盖）、 澜起科技。4）AIDC：东阳光、润泽科技（已覆盖）。 风险提示：下游需求不及预期；RD ...

当全球算力竞赛步入"万卡级"甚至"十万卡级"时代，大模型训练的效率瓶颈正悄然从芯片算力转向网络 互联。 3月12日，中科曙光在郑州正式发布全栈自研的400G无损高速网络产品scaleFabric，成为国内首个在高 端RDMA领域实现技术突破的厂商。从底层112G SerDes IP到上层管理软件，该系统均为100%自研产 品，端到端时延低至0.9微秒，单子网互连规模达传统InfiniBand的2.33倍，理论最大支持11.4万卡集群 部署。 更具说服力的是，这套国产网络已在国家超算互联网位于郑州的核心节点稳定运行超10个月，支撑起3 万卡规模的智算集群，承载真实大模型训练任务。 中科曙光高级副总裁李斌在会后对21世纪经济报道等媒体表示："从万卡到十万卡，最核心的技术突破 不在计算节点，而在互联系统。" 这标志着，中国在智算基础设施的关键一环——高速网络领域，已从"跟跑"走向"并跑"，并试图以开放 生态解构英伟达的封闭生态。 为何网络成算力关键瓶颈？ 大模型训练对算力的需求早已不是秘密，但一个容易被忽视的事实是：当集群规模从千卡扩展到万卡甚 至十万卡，计算节点之间的通信效率，正成为决定整体算力利用率的关键变量。 ...

Core Viewpoint - The article highlights the significant technological advancement made by Zhongke Shuguang in the high-end RDMA network sector, marking a shift from "catching up" to "running alongside" in China's intelligent computing infrastructure, particularly in high-speed networking [1][7]. Group 1: Technological Breakthroughs - Zhongke Shuguang has launched the scaleFabric, a fully self-developed 400G lossless high-speed network product, achieving a low end-to-end latency of 0.9 microseconds and supporting a theoretical maximum of 114,000 card cluster deployments [1][3]. - The scaleFabric system has been operational for over 10 months at the National Supercomputing Internet core node in Zhengzhou, successfully supporting a 30,000 card intelligent computing cluster [6][10]. - The product's core technology indicators include a link fault recovery time of less than 1 millisecond and a single subnet interconnection scale 2.33 times that of traditional InfiniBand [3][4]. Group 2: Market Dynamics - The demand for high-speed networks is driven by the increasing number of network ports required as cluster sizes grow from thousands to tens of thousands of cards, leading to a 10 to 20 times increase in network usage [3][10]. - The current market is dominated by NVIDIA, particularly through its InfiniBand technology, which has established a closed-loop ecosystem that poses a challenge for competitors [7][10]. - Zhongke Shuguang aims to provide a more open and adaptable network solution that can integrate with various domestic computing chips, thereby reducing dependency on a single vendor [10][12]. Group 3: Strategic Positioning - Zhongke Shuguang's strategy involves inheriting the open aspects of InfiniBand while attempting to break the commercial binding within NVIDIA's ecosystem [9][12]. - The company plans to establish a working group under the Guanghe organization to create relevant standards and enhance ecosystem adaptability through user feedback [9][12]. - The anticipated market for high-speed networks is substantial, with a focus on scientific engineering computing and AI training, indicating numerous potential deployment opportunities [12][10].

这并非一次简单的产品迭代，而是国产算力基础设施从 " 堆卡数量 " 迈向 " 网络效率 " 的关键一 跃。当算力堆到一定程度，真正决定集群价值的，早已不是单点芯片的性能，而是让算力 " 跑起来 " 的大动脉 —— 算力的下半场，比的就是这张 " 网 " 。 据 国家超算互联网平台披露，核心节点试运行邀测以来已吸引数百家 AI 模型厂商及诸多高校和科研 院所参与。 在 AI 时代，算力被视为支撑一切的基础。一个大规模智算中心 的迅速投运是当下各行各业如火如 荼扩大 AI 应用的缩影。更不易见但更关键的变化藏在这些 大规模集群算力基础设施 的建设中。 中 国信息通信研究院云计算与大数据研究所云计算部副主任郑立介绍，超大规模智算集群服务是全球 AI竞争的关键，国内企业正积极开展融合与自研的解决方案。 一个月前， scaleFabric 高速网络已经部署在国家超算互联网核心节点，与 3 套曙光万卡超集群、超 3 万张国产 AI 加速卡一同编织成一张高效的算力网。 随着大模型训练规模迈向万卡级，智算基础设 施正进入"网络定义算力效率"的新赛点。3月12日，"scaleFabric高速网络产品发布会"正式召开，中 科曙光 ...

Group 1 - The core viewpoint of the article highlights the breakthrough in domestic RDMA technology with the launch of the scaleFabric, a native lossless RDMA high-speed network by Zhongke Shuguang, which competes with Nvidia's NDR technology [2][4] - The scaleFabric400 series network products have been validated in a nearly 10,000-card scale environment and have been running stably for over 10 months, filling a technological gap in domestic high-speed interconnects for clusters [2][8] - The performance specifications of the scaleFabric400 include a port bandwidth of 400Gbps, end-to-end communication latency as low as 0.9 microseconds, and a switch with a single port bandwidth of 800Gbps, supporting a total switching capacity of 64Tbps [6][8] Group 2 - The product features a credit-based lossless flow control mechanism that mitigates congestion and packet loss risks, with a link failure recovery time of less than 1 millisecond, supporting nearly 10,000-card clusters [8][11] - Compared to Nvidia's NDR, the scaleFabric400 offers a 25% increase in switch port density, a 100% increase in maximum QP number supported by network cards, and a maximum interconnect scale that is 2.33 times that of traditional IB [8][10] - The deployment of the scaleFabric network is already operational in Zhengzhou, supporting a national-level AI computing network base with a total scale of 30,000 cards [9][10] Group 3 - Zhongke Shuguang has developed a complete computing power foundation through long-term technological accumulation in high-performance computing, storage, and networking, enabling a collaborative development of "computing-storage-network" [13] - The successful implementation of the native RDMA network signifies the formation of an independent technological path in intelligent computing interconnects in China, addressing a critical component of the country's computing infrastructure [13] - The high-performance network industry ecosystem surrounding the native RDMA technology is accelerating its formation as the product is applied in ultra-large-scale intelligent computing clusters [13]

Core Insights - The A-share technology sector experienced a notable pullback, with the AI ETF (515070) dropping over 1.3% during trading, while holdings like Unisplendour surged over 4% [1] - Nvidia's CEO Jensen Huang showcased the company's full-stack autonomous driving software platform, DRIVE AV, during a 22-minute video ride in an autonomous vehicle, highlighting its capabilities without human intervention [1] - Huatai Securities believes the AI industry is in a rapid development phase, with expanding technological innovations and application scenarios, indicating significant long-term investment value in the AI sector [1] Industry Summary - The AI ETF (515070) tracks the CS AI Theme Index (930713), selecting stocks that provide technology, foundational resources, and applications in the AI sector, focusing on the midstream and upstream of the AI industry chain [1] - Key weight stocks in the ETF include major domestic technology leaders such as Zhongji Xuchuang, Xinyisheng, Cambricon Technologies, and Hikvision, among others [1] - The AI industry is expected to benefit from increased policy support and market demand, particularly in foundational technology areas like algorithm frameworks, which are seen as future investment focal points [1]

Investment Rating - The report maintains a positive investment rating for the thermal management materials industry [3]. Core Insights - The rapid growth of high-power chips due to AIGC has led to a significant increase in thermal flow density, making liquid cooling the preferred solution for high-power chip cooling [6][23]. - The demand for liquid cooling solutions is expected to rise, driven by advancements in AI data centers and the increasing power requirements of data center AI cabinets [6][50]. - The report identifies three main investment opportunities in liquid cooling: high-power chips from NVIDIA, liquid cooling in high-power data centers, and thermal management for electric vehicle batteries and energy storage [6]. Summary by Sections Current Mainstream Cooling Solutions - Cooling methods are categorized into active and passive cooling, with active cooling including liquid cooling and air cooling [8][9]. - Liquid cooling has become the mainstream solution for high-performance scenarios, particularly in AI servers and data centers [8][9]. Thermal Management Industry Chain - The thermal management industry chain includes upstream raw materials, midstream manufacturing, and downstream applications, with the midstream segment being the most competitive [14][18]. - Key raw materials include high-purity copper and aluminum alloys, which are expected to see increased demand due to the rise in liquid cooling applications [6][11]. Domestic and Foreign Competition - The report highlights the competitive landscape, noting that foreign companies dominate high-end thermal management materials, while domestic companies are catching up and have significant room for growth [19][22]. Liquid Cooling Market Growth - The global market for liquid cooling plates in power batteries is projected to reach 14.5 billion yuan by 2025, with domestic market size expected to reach 9.6 billion yuan [6]. - The growth of energy storage batteries is also anticipated to drive demand for liquid cooling solutions [6]. Key Companies and Their Prospects - Major companies benefiting from the liquid cooling trend include Cooler Master, Vertiv, and domestic firms like Yingwei and Highlan [6][42]. - Companies like Jetcool are developing innovative solutions for high-power chips, which are expected to enhance overall efficiency in AI clusters [40]. Future Trends and Developments - The report emphasizes the importance of liquid cooling in meeting the energy efficiency requirements set by regulatory bodies for data centers [46][50]. - The shift towards liquid cooling is driven by the need to manage increasing thermal loads effectively, as traditional air cooling approaches reach their limits [50][51].