如何看光模块未来几年增长空间？20260108 摘要 光模块行业受益于数据中心升级和 AI 大模型需求，800G 光模块增量主 要来自 AI 应用。SKU 层互联需求增加，高速光模块需求呈指数级增长， 预计 2025-2028 年行业保持高增长。 海外大厂如谷歌、英伟达采用 Skyop 设计，增加光连接需求。谷歌预计 2026 年展示机柜内光互联方案，英伟达或稍晚。高密度机柜内部及 Skill Up 层之间将大量使用高速光模块。 硅光集成方案、LPO 等技术路径市场份额将提升。2025 年硅光在单模 光模块中占比 20%-30%，预计 2026 年翻倍至 40%-50%，主要因 800G 硅光产品成熟及 100G EML 芯片紧缺。 LPO 适用于小应用推理场景，LRO 和 TRO 通过定制化设计降低整体链 路功耗，适用于特定连接场景。这些定制化互联解决方案将在未来批量 应用于特定连接场景。 英伟达和博通积极推动 CPO 方案，但共封装结构导致维修复杂且成本高 昂，限制大规模应用。NPU 通过集成光模块缩短电信号传输距离，头部 厂商已开始研发。 Q&A 2026 年和 2027 年光模块行业的增长趋势如何？ ...

今天分享的是：算力专题：全球算力十大趋势2026 报告共计：62页 《全球算力十大趋势2026》报告指出，算力已成为驱动数字经济与智能变革的核心引擎，全球算力规模呈指数级增长，各国纷纷将其上升为国家战略，科技 巨头持续加码布局，算力竞赛白热化。AI大模型正演进为智能世界的"底层操作系统"，加速渗透千行百业，从技术验证迈向商业闭环，带动算力需求爆发式 增长，而智能体的复杂任务处理进一步倒逼算力供给体系升级。数字智能正向具身智能演进，世界模型作为核心支撑打破虚实边界，推动智能体在物理世界 实现深度交互，开辟万亿级市场空间。技术架构层面，传统以CPU为中心的模式逐渐被CPU、GPU、NPU等多样化平等架构取代，超节点成为智算中心新底 座，通过高速互联突破"内存墙"瓶颈，支撑超大规模模型高效运行。算力与网络深度融合，毫秒级算力网建设梯次推进，实现"算力随需即取"，超算与智算 走向深度融合，构筑科学计算新范式。开源开放成为算力生态核心，头部厂商通过全栈开源降低创新门槛，推动产业协同发展。智算中心向高密化、液冷 化、集群化演进，绿电直供模式破解能耗难题，量子计算则进入工程化关键期，未来1-2年将迎来从技术突破到商业化应用 ...

Core Viewpoint - The article highlights the significant role of NVIDIA in the AI boom, attributing its success not only to GPUs but also to its strategic acquisition of Mellanox, which has greatly enhanced its networking capabilities. This has led to a substantial increase in networking revenue, showcasing the growing importance of networking in the AI era [1]. Group 1: NVIDIA's Success and DPU's Role - NVIDIA's networking revenue grew by 162% year-on-year to $8.2 billion in Q3 2025, surpassing the $6.9 billion paid for Mellanox [1]. - The emergence of DPU (Data Processing Unit) has become crucial in modern data centers, as it offloads tasks from CPUs, enhancing overall system performance [2][3]. - DPU is seen as a key component in creating a secure and accelerated data center, integrating CPU, GPU, and DPU into a single programmable unit [2]. Group 2: DeepSeek's Insights on DPU - DeepSeek emphasizes the importance of DPU in AI infrastructure, suggesting that integrated communication co-processors in DPUs could be vital for next-generation AI hardware [4]. - The use of RDMA (Remote Direct Memory Access) in DPU enhances online inference throughput and computational efficiency by minimizing resource contention [5]. Group 3: Cloud Leopard Technology's Breakthrough - Cloud Leopard Technology has successfully produced China's first 400Gbps DPU chip, achieving global top-tier performance with capabilities to process millions of data packets per second and low latency of 5 microseconds [8][10]. - The company has gained recognition from major investors and has been able to produce complex chips without modifying any transistors, demonstrating its technological prowess [7][8]. - Cloud Leopard aims to launch an 800Gbps network card to compete with NVIDIA's CX8 network card, further solidifying its position in the market [13]. Group 4: Industry Trends and Future Outlook - The article notes that various chip sectors, including CPU, GPU, and AI computing chips, have seen significant advancements and IPOs, indicating a fruitful period for the domestic chip industry [15]. - Cloud Leopard is positioned to potentially become the "first DPU stock in China," reflecting its growing influence in the semiconductor landscape [15].

Core Viewpoint - The embedded computing world is undergoing a transformation where AI is reshaping the architecture of MCUs, moving from traditional designs to those that natively support AI workloads while maintaining reliability and low power consumption [2][5]. Group 1: MCU Evolution - The integration of NPU in MCUs is driven by the need for real-time control and stability in embedded systems, particularly in industrial and automotive applications [3][4]. - NPU allows for "compute isolation," enabling AI inference to run independently from the main control tasks, thus preserving real-time performance [3][5]. - Current edge AI applications typically utilize lightweight neural network models, making hundreds of GOPS sufficient for processing, which contrasts with the high TOPS requirements in mobile and server environments [5]. Group 2: Major MCU Players' Strategies - TI focuses on deep integration of NPU capabilities in real-time control applications, enhancing safety and reliability in industrial and automotive scenarios [7][8]. - Infineon leverages the Arm ecosystem to create a low-power AI MCU platform, aiming to reduce development barriers for edge AI applications across various sectors [9][10]. - NXP emphasizes hardware scalability and a full-stack software approach with its eIQ Neutron NPU, targeting diverse neural network models while ensuring low power and real-time response [11][12]. - ST aims for high-performance edge visual applications with its self-developed NPU, pushing the boundaries of traditional MCU AI capabilities [13][14]. - Renesas combines high-performance cores with dedicated NPU and security features, focusing on reliable edge AIoT applications [15][16]. Group 3: New Storage Technologies - The introduction of NPU in MCUs necessitates a shift from traditional Flash storage to new storage technologies that can handle the demands of AI workloads and frequent updates [17][18]. - New storage solutions like MRAM, RRAM, PCM, and FRAM are emerging to address the limitations of Flash, offering advantages in reliability, speed, and endurance [21][22][25][28][30]. - MRAM is particularly suited for automotive and industrial applications due to its high reliability and endurance, with companies like NXP and Renesas leading in its adoption [22][23][24]. - RRAM offers benefits in speed and flexibility, making it a strong candidate for AI applications, with Infineon actively promoting its integration into next-generation MCUs [25][26][27]. - PCM provides high storage density and efficiency, suitable for complex embedded systems, with ST advocating for its use in advanced MCU designs [28][29]. Group 4: Future Implications - The dominance of Flash storage is being challenged as new storage technologies demonstrate superior performance and reliability for embedded systems [33]. - The integration of NPU and new storage technologies in MCUs represents a shift towards system-level optimization, enhancing overall performance and efficiency [33]. - The transformation in the MCU market presents structural opportunities for domestic manufacturers to innovate and compete against established international players [33].

在《MPW零度对话》系列中，我们邀请中国最具知名度和影响力的女性领导者，讨论当下所有 人共同关心的话题，从充满热情和妙趣的对话中，提炼出冷静理性的智慧。 临近年末，多家权威词典公布的年度词汇均指向 AI，例如"slop"（网络垃圾）、"vibe coding"（氛围编 程）与"rage bait"（愤怒诱饵）。而在中国，几家机构联合推选的年度国内词则是"DeepSeek"。 泡沫之外，张璐更加关注的是今年AI领域呈现的新趋势。在她看来，2025年AI领域经历了"上升"和"下 沉"——一边是全球竞赛中的技术持续突破，一边是AI加速向产业深处落地。 技术"上升" 在AI基础设施层，尤其是芯片领域，过去由GPU主导的格局正逐步走向多元。张璐观察到，一些新模 型架构在CPU上运行效率更高；谷歌的TPU发展迅猛；高通、英特尔推出的NPU则在能效方面表现突 出。 在云基础设施层面，长期困扰行业的四大难题——算力成本高、能耗大、边缘设备应用难、数据隐私问 题——正逐步得到解决。例如，OpenAI的token价格已从每千个30美元大幅降至9美分；被英伟达收购的 由华人创立的Lepton公司，其技术能显著降低GPU消耗。 能 ...

智通财经APP获悉，东吴证券发布研报称，2026年AI算力产业链预计将迎来业绩放量。云端算力方面， 国产GPU进入业绩兑现期，产业链协同至关重要。端侧算力则受益于AI终端创新与3D DRAM放量，多 场景加速落地。同时，先进制程晶圆代工扩产、HVDC电源架构升级以及光铜互联需求增长，共同驱动 产业链各环节技术升级与价值提升。 东吴证券主要观点如下： 云端算力：迎接国产算力产业链上下游共振带动业绩放量 3DDRAM：端侧AI存储26年为放量元年，产业趋势逐渐确立 26年AI硬件落地带来存力需求的快速提升，高带宽/低成本的3DDRAM有望在多领域放量。该行认为机 器人/AIOT/汽车等领域对本地大模型的部署离不开3DDRAM存储的支持，其为各端侧应用从"能 用"到"好用"的关键硬件革新。多款NPU的流片发布亦为3DDRAM提供丰富适配场景。此外，手机/云端 推理等场景亦将逐步导入，成为26H2及27年关键场景，应用领域持续拓宽。相关公司兆易创新、北京 君正等。 端侧AI模型：架构优化突破物理瓶颈，利益分润决定生态版图 2026年国产算力芯片龙头有望进入业绩兑现期，看好国产GPU受益于先进制程扩产带来的产能释放。考 ...

Investment Rating - The report maintains a rating of "Buy" for the electronic industry [1] Core Insights - The electronic industry is expected to experience significant growth driven by advancements in AI technology and the domestic supply chain's maturation, particularly in cloud and edge computing [11][15] - The report highlights the importance of 3D DRAM as a key hardware innovation for AI applications, with expectations for substantial demand growth in 2026 [22][27] - The shift towards high-density interconnects and advanced power supply architectures is crucial for supporting the increasing power density of AI data centers [50][56] Summary by Sections Cloud Computing Power - The domestic computing power supply chain is accelerating, with significant performance releases expected from local manufacturers like Zhongke Shuguang and Huawei [11] - The transition from Scale-Out to Scale-Up networking is enhancing bandwidth and reducing latency, which is critical for AI applications [11] Edge Computing Power - The integration of edge and cloud computing is becoming essential for AI applications, with edge devices benefiting from advancements in SoC technology [15][17] - Companies like Jingchen and Ruixinwei are positioned to capitalize on the growing demand for edge AI solutions [19] 3D DRAM - 3D DRAM is anticipated to become mainstream in 2026, driven by its high bandwidth and low cost, making it essential for various AI applications [22][27] - Companies such as Zhaoyi Innovation are expected to lead in the development of 3D DRAM technologies [28] AI Models - The optimization of AI models is crucial for enhancing performance and user experience, with a focus on local processing capabilities [29][30] - The collaboration between terminal manufacturers and model providers is expected to evolve, shaping the competitive landscape [30][33] AI Terminals - 2026 marks the beginning of a new era for AI terminals, with major companies like Meta, Apple, and Google launching innovative products [34][36] - The development of new terminal forms, such as smart glasses and desktop robots, is expected to drive market growth [34][35] Longxin Chain - Longxin's expansion plans are set to enhance the DRAM supply chain, with a focus on 3D architecture to improve performance and efficiency [38][39] - The company is expected to benefit from increased capital investment and technological advancements [39][41] Wafer Foundry - The domestic wafer foundry industry is entering a new expansion phase, particularly in advanced logic processes [42][43] - Key players like SMIC and Huahong are expected to lead this expansion, addressing the growing demand for advanced chips [44] PCB Industry - The PCB market is poised for growth, driven by the demand for high-performance materials and advanced designs [45][48] - Companies like Shenghong Technology are expected to benefit from the rising demand for AI-related PCB applications [49] Optical-Copper Interconnection - The demand for optical and copper interconnections is increasing, driven by the growth of AI computing clusters [50][52] - Companies such as Changguang Huaxin are well-positioned to capitalize on this trend [53] Server Power Supply - The shift to HVDC power supply architectures is becoming essential for AI data centers, addressing the challenges of increasing power density [55][56] - Companies like Oulu Tong are expected to lead in the development of advanced power supply solutions [56]

（文章来源：证券日报） 证券日报网讯 12月30日，星宸科技在互动平台回答投资者提问时表示，公司有自研的NPU，搭配可拓 展的算力架构和分布式算力芯片组，最高可覆盖至128T算力（如面向具身智能机器人大小脑等高端智 能场景）。目前暂未涉及TPU。 ...

(原标题：星宸科技(301536.SZ)：目前暂未涉及TPU) 格隆汇12月30日丨星宸科技(301.536.SZ)在互动平台表示，公司有自研的NPU，搭配可拓展的算力架构和分布式算力芯片组，最高可覆盖至128T算 力（如面向具身智能机器人大小脑等高端智能场景）。目前暂未涉及TPU。 相关ETF 食品饮料ETF (产品代码: 515170) ★ 跟踪：中证细分食品饮料产业主题指数 近五日涨跌: -1.42% 市盈率： 19.99倍 资金流向: 最新份额为103.8亿份,减 少了1.8亿份,净申赎-1.0亿 元。 估值分位:17.29% 游戏ETF (产品代码：159869) 资金流向: 最新份额为87.0亿份,增加 了5800.0万份,净申赎 8382.7万元。 估值分位:56.54% 科创50ETF (产品代码: 588000) ★ 跟踪：上证科创板50成份指数 近五日涨跌:0.85% 市盈率:160.92倍 资金流向: 最新份额为543.4亿份,增 加了4.9亿份，净申赎7.0亿 元。 ★ 跟踪：中证动漫游戏指数 近五日涨跌:0.48% 市盈率： 37.50倍 云计算50ETF (产品代码：516630) ...

Core Insights - A Chinese research team has developed a new type of chip based on resistive random-access memory (RRAM) that achieves a precision of 24-bit fixed-point accuracy in analog matrix computations, marking a significant advancement in computational efficiency and energy consumption for AI applications [2][12][15] - This chip can support various cutting-edge applications, including 6G communication, embodied intelligence, and AI model training, while being produced using mature 28nm technology, thus avoiding reliance on advanced lithography processes [2][4][10] Technology Overview - The new chip represents a departure from traditional digital computing paradigms, which rely on binary logic and silicon-based transistors, to a more efficient analog computing approach that directly utilizes physical laws for calculations [4][6][15] - The precision of analog computing has been significantly improved, reducing relative error from 1% to one part in ten million (10⁻⁷), which is crucial for large-scale computations where errors can accumulate exponentially [8][12][15] Innovation Highlights - The chip's innovations include the use of RRAM as a core component, a novel feedback circuit design that minimizes energy consumption while enhancing accuracy, and the implementation of classic iterative optimization algorithms for efficient matrix equation solving [15][16] - The chip's architecture allows for high-speed, low-power solutions to matrix equations, making it suitable for applications that require rapid computations, such as second-order training methods in AI [19][21] Application Potential - The chip is particularly well-suited for medium-scale applications, such as AI model training and 6G MIMO systems, where it can outperform traditional digital chips [18][25] - Future plans include scaling the chip's matrix size from 16x16 to 128x128 within two years, with aspirations to reach 512x512, which would enhance its applicability in various computational scenarios [25][26] Strategic Value - This development provides China with a potential alternative to reliance on advanced processes and NVIDIA GPUs, positioning the country favorably in the global computational landscape [10][11] - The successful demonstration of this new computing paradigm is seen as a critical step towards addressing future computational demands, emphasizing the need for ongoing investment in technology and infrastructure [11][26]