截至2025年9月30日 10:11，中证半导体产业指数上涨0.92%，成分股神工股份上涨6.19%，芯源微上涨5.13%，盛美上海上涨4.69%，京仪装备上涨3.83%，晶 升股份上涨3.81%。半导体产业ETF(159582)上涨1.14%，最新价报2.32元。拉长时间看，截至2025年9月29日，半导体产业ETF近1周累计上涨12.15%，涨幅 排名可比基金1/2。 流动性方面，半导体产业ETF盘中换手10.67%，成交3743.92万元，市场交投活跃。拉长时间看，截至9月29日，半导体产业ETF近1周日均成交9081.04万 元。 1） OpenAI 宣布将于2025年10月6日在美国旧金山举行第三届年度开发者大会，主题聚焦AI技术在硬件领域的应用拓展，预计吸引超过1500名开发者参与。 大会将展示公司正在研发的多款硬件产品，包括无显示屏智能设备、智能眼镜和可穿戴设备等，这些产品均深度集成ChatGPT大模型或自研AI芯片。 截至2025年9月30日 10:11，上证科创板芯片指数强势上涨1.80%，成分股佰维存储上涨11.31%，燕东微上涨5.49%，芯源微上涨5.13%，源杰科技，盛美上海 等个股 ...

Group 1 - The demand for lower computing costs and diversified supply chain risks is driving the performance surge of overseas ASIC chip giants like Broadcom and Marvell, while domestic ASIC chip companies are also experiencing a significant increase in orders [1] - Chipone Technology (688521.SH), known as "China's first semiconductor IP stock," reported new orders of 1.205 billion yuan from July 1 to September 11, marking an 85.88% increase compared to the entire third quarter of 2024, with AI computing-related orders accounting for approximately 64% [1] - The AI computing-related orders primarily refer to ASIC chip design services, catering to customized chip demands from chip design companies, internet firms, and cloud service providers [1] Group 2 - A cost comparison by Southwest Securities shows that the unit computing costs of Google's fifth-generation TPU and Amazon's Trainium 2 ASIC chips are 70% and 60% of NVIDIA's H100 chip, respectively [2] - General-purpose GPUs are favored for model training due to their versatility, while ASIC chips are more efficient for specific tasks, leading to a fragmented market where NVIDIA dominates model training but faces competition in model inference from ASIC players [2][3] - NVIDIA is responding to the competition by releasing a new chip designed specifically for AI inference, aimed at improving cost-effectiveness by reducing unnecessary high-cost configurations [2] Group 3 - Chipone Technology, founded in 2001, has established itself as a leading ASIC chip service provider in China, with a comprehensive IP system that includes various types of processors and over 1,600 mixed-signal and RF IPs [3] - The demand for AI ASICs is surging due to the large-scale deployment of large models, with the Chinese AI chip market projected to reach 142.537 billion yuan in 2024, where GPU chips will hold approximately 69.9% of the market share [3] Group 4 - In the first half of 2025, AI computing-related revenue is expected to account for about 52% of Chipone Technology's chip design business [4] - Traditional general-purpose GPUs are increasingly unable to meet the specific demands of certain scenarios, while AI ASICs offer high cost-effectiveness and low power consumption due to their customized architecture [5] Group 5 - Chipone Technology is seeking to acquire RISC-V architecture CPU IP provider Chipone Technology, which is expected to enhance its AI ASIC business [5] - The company relies on a partnership with UK IP giant Alphawave for high-speed SerDes IP, which is crucial for high-speed data transmission [5] Group 6 - The domestic AI ASIC landscape includes major players like Huawei, Alibaba, and Baidu, with products such as Huawei's Ascend series and Alibaba's Pingtouge [6] - The rapid expansion of domestic AI chips is driven by technological breakthroughs from large internet companies and local suppliers [8] Group 7 - The global AI ASIC market is projected to grow from approximately $6.6 billion in 2023 to $55.4 billion by 2028, with a compound annual growth rate of 53% [9] - Major cloud providers like Google and Amazon are leading the self-developed ASIC chip trend, significantly boosting the performance of ASIC service providers [10] Group 8 - Broadcom's AI business reported $5.2 billion in revenue for the third quarter of 2025, a 63% year-on-year increase, with a new major customer ordering over $10 billion in custom AI chips [10] - The competitive landscape is shifting, with concerns that NVIDIA's clients may pivot from GPUs to ASICs as the latter gain traction [11] Group 9 - Despite NVIDIA's skepticism about the flexibility of ASICs, the company is actively developing new GPU architectures to compete in the inference market [12][18] - The coexistence of ASICs and general-purpose GPUs is expected, with each technology serving different application scenarios effectively [18]

Core Viewpoint - The article highlights the significant advancements in China's semiconductor industry, particularly the development and potential market impact of domestically produced HBM3 memory chips, which could disrupt the current dominance of major global players like Samsung, SK Hynix, and Micron [2][5]. Group 1: HBM3 Development and Market Impact - Domestic DRAM leader has begun supplying HBM3 samples to Huawei, manufactured using a self-developed 16nm G4 process, awaiting mass production approval [2]. - The G4 process allows for a 20% reduction in chip size and significant energy efficiency improvements compared to the previous 18nm G3 process, positioning China as the third country capable of mass-producing HBM3 after South Korea and the USA [2]. - The integration of HBM3 chips into Huawei's Ascend 910C AI chip is expected to enhance AI computing capabilities significantly, with a 40% increase in inference speed and a 30% improvement in energy efficiency [3][4]. Group 2: Competitive Landscape - The global HBM market is currently dominated by SK Hynix, which holds over 50% market share, followed closely by Samsung and Micron [4]. - In response to the emergence of domestic HBM3, international competitors are accelerating their technology iterations, with SK Hynix planning to start mass production of the world's first 12-layer HBM3E by September 2024 [4]. - Micron aims to achieve parity in HBM market share with its overall DRAM share (approximately 25%) by the second half of 2025 [4]. Group 3: Future Outlook and Strategic Positioning - Analysts predict that the large-scale application of domestic HBM3 will compel international manufacturers to accelerate technology transfer, potentially leading to a 20%-30% decrease in HBM3e prices over the next two years [5]. - The Chinese semiconductor industry is leveraging "cost-performance + localization" strategies to capture a share of the mid-to-high-end market [5]. - By 2025, domestic HBM market demand is expected to exceed 120 million GB, accounting for 30% of the global total, driven by policies such as the "East Data West Computing" project [6].

Core Viewpoint - Huawei's CloudMatrix 384 represents a next-generation AI data center architecture designed to meet the increasing demands of large-scale AI workloads, featuring a fully interconnected hardware design that integrates 384 Ascend 910C NPUs and 192 Kunpeng CPUs, facilitating dynamic resource pooling and efficient memory management [6][55]. Summary by Sections Introduction to CloudMatrix - CloudMatrix is introduced as a new AI data center architecture aimed at reshaping AI infrastructure, with CloudMatrix 384 being its first production-level implementation optimized for large-scale AI workloads [6][55]. Features of CloudMatrix 384 - CloudMatrix 384 is characterized by high density, speed, and efficiency, achieved through comprehensive architectural innovations that lead to superior performance in computing, interconnect bandwidth, and memory bandwidth [2][3]. - The architecture allows for direct full-node communication via a unified bus (UB), enabling dynamic pooling and unified access to computing, memory, and network resources, which is particularly beneficial for communication-intensive operations [3][7]. Architectural Innovations - The architecture supports four foundational capabilities: scalable communication for tensor and expert parallelism, flexible heterogeneous workload resource combinations, a unified infrastructure for mixed workloads, and memory-level storage through decomposed memory pools [8][9][10]. Hardware Components - The core of CloudMatrix 384 is the Ascend 910C chip, which features a dual-chip package providing a total throughput of up to 752 TFLOPS and high memory bandwidth [17][18]. - Each computing node integrates multiple NPUs and CPUs, connected through a high-bandwidth UB network, ensuring low latency and high performance [22][24]. Software Stack - Huawei has developed a comprehensive software ecosystem for the Ascend NPUs, known as CANN, which facilitates efficient integration with major AI frameworks like PyTorch and TensorFlow [27][33]. Future Directions - Future enhancements for CloudMatrix 384 include integrating VPC and RDMA networks, expanding to larger supernode configurations, and pursuing finer-grained resource decomposition and pooling [58]. - The architecture is expected to evolve to support increasingly diverse AI workloads, including specialized accelerators for various tasks, enhancing flexibility and efficiency [47][48]. Performance Evaluation - CloudMatrix-Infer, a service solution built on CloudMatrix 384, has demonstrated exceptional throughput and low latency in processing tokens during inference, outperforming leading frameworks [57]. Conclusion - Overall, Huawei's CloudMatrix is positioned as an efficient, scalable, and performance-optimized platform for deploying large-scale AI workloads, setting a benchmark for future AI data center infrastructures [55][58].

Group 1: Core Views - The communication sector outperformed the market in May, with the communication index rising by 3.98% compared to a 1.85% increase in the CSI 300 index, ranking 9th among 31 primary industries [1][14] - The PE valuation of the communication industry is showing a recovery trend, currently at 20.9 times, which is relatively low compared to historical levels [18][22] - Key sub-sectors such as military informationization, optical modules, and satellite internet have performed well, with respective increases of 11.2%, 10.6%, and 8.1% [1][26] Group 2: AI Investment Trends - Major internet companies like Alibaba and Tencent are significantly increasing their AI investments, with Alibaba's capital expenditure reaching 24.612 billion yuan, up 120.68% year-on-year, and Tencent's at 27.476 billion yuan, up 91.35% year-on-year [30][39] - Nvidia reported a 69% year-on-year revenue growth in Q1 2026, driven by a 73% increase in data center revenue, indicating sustained high demand for AI [53][58] - CoreWeave, a GPU cloud service provider, reported a 420% year-on-year revenue growth in Q1 2025, highlighting the booming demand for GPU as a service [73] Group 3: Commercial Aerospace Development - China's private rocket sector is accelerating, with multiple commercial launch sites being planned, including four in Hainan and four in Jiuquan [3] - The successful flight recovery test of the Yuanxing-1 rocket by Arrow Yuan Technology marks a significant milestone in China's commercial space endeavors [3] Group 4: Investment Recommendations - The report suggests focusing on AI cloud-side and edge-side developments, as well as changes in commercial aerospace, recommending stocks like China Mobile, Tianfu Communication, and Huagong Technology for long-term investment [4][5] - The three major telecom operators are noted for their stable operations and increasing dividend ratios, making them attractive for long-term investment [4]

Core Viewpoint - Malaysia is enhancing its AI sovereignty by deploying Ascend GPUs and localizing servers with models like Deepseek, which is expected to accelerate the overseas expansion of domestic computing power hardware and software architectures [1][2]. Group 1: Deployment of Ascend Chips - Malaysia's Ministry of Communications announced the advancement of its AI infrastructure strategy, supported by Ascend GPUs and the Deepseek model [2]. - The first sovereign generative AI server, AlterMatic DT250AI, outperforms the industry average by 20% and is already adopted by various government agencies [2]. - Skyvast and Liyang Chips plan to deploy 3,000 advanced GPUs across multiple infrastructure areas by 2026, supported by Malaysia's AI system integration ecosystem [2]. Group 2: Hardware and Software Architecture Upgrades - The Ascend 910C single card has a BF16 computing power of approximately 780 TFLOPS, nearing 80% of the H100's performance [3]. - The CloudMatrix 384 super node can achieve a single card decode throughput of 1920 tokens/s under a single user 20 TPS condition, comparable to H100 performance [3]. - MindSpore 2.6 fully supports the training and inference processes for models like Deepseek V3/R1 MoE, enhancing usability for mainstream SOTA models [3]. Group 3: Catalysts for Growth - The performance upgrades of domestic computing power chips and breakthroughs in advanced manufacturing processes are key catalysts for growth in the sector [4].

以下文章来源于豹变 ，作者朱晓宇 豹变 . 直抵核心。做最具穿透力、洞察力的商业观察，深度影响未来。 来源丨豹变（ID：baobiannews） 作者丨朱晓宇 编辑丨邢昀 图源丨Midjourney 时隔三个月，英伟达创始人黄仁勋紧急访华。与今年1月春节期间的常规行程有所不同，在对等关税的进 一步升级下，英伟达针对中国大陆市场推出的H20芯片也被进一步限制出口，股价暴跌，黄仁勋急需找到 中国市场的突破口。 重要性之高，连黄仁勋最爱的皮衣都换成了西装。黄仁勋在会谈中表示，中国是英伟达非常重要的市 场，希望继续与中国合作。 英伟达作为全球AI算力芯片的龙头，在美国禁令发布后遭受重创，黄仁勋的紧急访华行程背后，折射出 全球芯片产业被改写的格局。 其中，中国加速国产芯片替代成为趋势。尤其是车规级芯片方面，曾经依赖美国进口芯片的上下游产业 链，正在紧急寻找更稳妥的替代方案，甩掉美国标签，更高比例的国产化芯片成为整车厂的核心诉求。 也正是在这轮操作的刺激下，国产车规级芯片正在迎来询价需求的大爆发。 国内一家生产高低边驱动控制器芯片的公司向《豹变》透露，美国的芯片出口管制给中国整车厂带来一 定冲击，4月开始，前来公司询价的 ...

Core Insights - The Chinese computing power market is experiencing a structural imbalance characterized by both surplus and shortage, with general computing power being relatively overabundant while intelligent computing power is in short supply [3][4][6]. Group 1: Market Dynamics - Several listed companies have announced winning bids for data center projects or signed computing power service contracts, indicating ongoing demand in the market [2][4]. - Major state-owned enterprises like China Mobile and China Telecom are significantly increasing their investments in computing power, with China Mobile planning a budget of 37.3 billion yuan for 2025, accounting for 25% of its total capital expenditure [5][4]. - The International Data Corporation (IDC) predicts that China's intelligent computing power will reach 1,037.3 EFLOPS by 2025 and 2,781.9 EFLOPS by 2028, reflecting a growing market scale [5][4]. Group 2: Structural Issues - The average rack utilization rate in China's IDC market is around 58%, indicating a significant amount of idle computing power [6]. - There is a notable disparity in computing power quality and regional distribution, with general computing power being overabundant in some areas while intelligent computing power is scarce, particularly in eastern regions where demand is high [6][7]. Group 3: Causes of Imbalance - The rapid growth and iteration of computing power demand, coupled with the transition from older to newer hardware, have led to mismatches in supply and demand [8]. - A lack of understanding among both buyers and sellers regarding the requirements and capabilities of intelligent computing centers has contributed to the imbalance [9][10]. - Some companies prioritize low costs in western regions for building computing centers, neglecting the necessary conditions for effective operation, which leads to mismatched resources [10]. Group 4: Future Outlook - The industry is expected to optimize and iterate on computing power scheduling, hardware, and supporting software to address the current challenges [11][14]. - The trend of "East Data West Computing" is emerging, where eastern data centers handle frequently accessed data while western centers manage less time-sensitive tasks [12]. - Domestic high-end computing hardware is accelerating in development, with companies like Huawei introducing competitive chips to fill the supply gap [13].

以下文章来源于豹变 ，作者朱晓宇 豹变 . 直抵核心。做最具穿透力、洞察力的商业观察，深度影响未来。 将投中网设为"星标⭐"，第一时间收获最新推送 国产芯片打响"抢位赛"。 作者丨 朱晓宇 编辑丨 邢昀 来源丨 豹变 时隔三个月，英伟达创始人 黄仁勋紧急访华。与今年 1 月春节期间的常规行程有所不同，在对等关税的进一步升级下，英伟达针对中国大陆市场推出 的 H 20 芯片也被进一步限制出口， 股价暴跌， 黄仁勋急需找到中国市场的突破口。 重要性之高，连黄仁勋最爱的皮衣都换成了西装。 黄仁勋在会谈中表示，中国是英伟达非常重要的市场，希望继续与中国合作。 英伟达作为全球 AI 算力芯片的龙头，在 美国禁令 发布后遭受重创，黄仁勋的紧急访华行程背后，折射出全球芯片产业被改写的格局。 其中， 中国加速国产 芯片 替代 成为趋势。尤其是车规级芯片方面，曾经依赖美国进口芯片的上下游产业链，正在紧急寻找更稳妥的替代方案，甩掉美 国标签，更高比例的国产化芯片成为整车厂的核心诉求。也正是在这轮操作的刺激下， 国产车规级芯片正在迎来询价需求的大爆发。 国内一家生产高低边驱动控制器芯片的公司向《豹变》透露，美国的芯片出口管制给中国 ...