2026.03. 05 本文字数：1514，阅读时长大约3分钟 作者 | 第一财经 郑栩彤 存储芯片方面，在存储缺货、涨价的背景下，国内厂商积极抓住市场机遇。据研究机构Omdia数 据，按产能和出货量计算，国内厂商长鑫科技已成为全球第四的DRAM（动态随机存取存储器）厂 商。技术研发方面，去年11月该公司推出DDR5和LPDDR5X移动端内存系列，称两个产品系列速 率、容量两个维度均位居业界第一梯队。另一家国产存储厂商长江存储也在扩产，武汉的长江存储三 期工地计划今年建成投产，预计带动上下游200家企业聚集。 半导体设备方面，新凯来去年3月在上海半导体设备展上首次亮相，展示了包括刻蚀产品、扩散产 品、薄膜产品及物理量测产品、X射线量测产品、光学量检测产品在内的6大类共31款半导体工艺和 检测装备，包括外延沉积设备EPI（峨眉山）、原子层沉积设备ALD（阿里山）、物理气相沉积设备 PVD（普陀山）等。 去年10月在湾区半导体芯片展上，新凯来子公司还发布了两款国产电子工程EDA（原理图和PCB） 设计软件以及新一代超高速实时示波器。其中，超高速实时示波器带宽突破90 GHz，将国产示波器 性能提升了500%，实现 ...

Core Viewpoint - The MTT S5000, developed by Moore Threads, is positioned as a competitive GPU for large model training and inference, showcasing performance that rivals international flagship products, marking a significant advancement in domestic computing power capabilities [1][3]. Group 1: MTT S5000 Performance - The MTT S5000 features a single card AI computing power of 1000 TFlops with liquid cooling and 920 TFlops with air cooling, alongside 80 GB of memory and a memory bandwidth of 1.6 TB/s [4]. - The S5000's performance has been reported to match or even exceed that of NVIDIA's H100 in certain multi-modal large model fine-tuning tasks [4][6]. - The architecture utilizes the fourth-generation MUSA architecture, optimized for large-scale AI training, and supports full precision calculations from FP8 to FP64 [6]. Group 2: Cluster Performance - The Kua'e Wan Card cluster built on the S5000 achieves a floating-point operation capability of 10 Exa-Flops, with an MFU of 60% in Dense model training and around 40% in MoE models, maintaining over 90% effective training time [8]. - The S5000 employs unique ACE technology for communication tasks, allowing for zero-conflict parallel computing and significantly enhancing model computing power utilization [10]. Group 3: Training and Inference Cases - In January 2026, the Zhiyuan Research Institute completed end-to-end training and alignment verification of the RoboBrain 2.5 model using a thousand-card cluster based on the S5000, achieving a training loss difference of only 0.62% compared to NVIDIA's H100 cluster [10]. - In December 2025, Moore Threads, in collaboration with Silicon-based Flow, conducted performance testing on the DeepSeek-V3 671B model using the S5000, achieving a record-breaking inference throughput of over 4000 tokens/s for Prefill and over 1000 tokens/s for Decode [12].

Core Insights - Moore Threads has officially open-sourced the TileLang-MUSA project, providing complete support for the TileLang programming language, aimed at lowering development barriers and enhancing AI and high-performance computing experiences on domestic computing platforms [2][4] Group 1: TileLang Overview - TileLang is a high-performance AI operator programming language based on tensor tiling abstraction, designed as a domain-specific language (DSL) with a declarative syntax similar to Python, allowing developers to describe computational intentions close to mathematical formulas [3] - The language significantly improves GPU computing development efficiency by abstracting complexities, enabling cross-platform capabilities, and automating complex tasks such as layout inference and memory optimization [3] Group 2: Practical Applications - The development of DeepSeek-V3 has utilized TileLang for rapid prototyping and performance validation, demonstrating its practical value in large-scale model training [4] - TileLang-MUSA serves as a bridge between advanced syntax and domestic computing, effectively unlocking the performance potential of full-featured GPUs [5] Group 3: Technical Features - TileLang-MUSA has been functionally validated on multiple generations of Moore Threads' full-featured GPUs, including the MTT S5000 and MTT S4000, showcasing good hardware compatibility [5] - The project has achieved over 80% unit test coverage for native TileLang operators based on the MUSA architecture, providing reliable assurance for large-scale applications [5] Group 4: Development Efficiency - Using TileLang-MUSA, the code volume has been reduced by approximately 90% compared to C++ code, resulting in clearer logic and significantly lower development and maintenance costs [6] Group 5: Future Outlook - The open-sourcing of TileLang-MUSA is a crucial step for Moore Threads in building a domestic computing ecosystem, with plans to create a unified acceleration platform covering everything from single operators to complete large models [7] - Future initiatives include performance optimization, integration with mainstream AI frameworks, and the development of debugging and performance analysis tools to support developers throughout the process [7]

Core Viewpoint - YunGongChang (02512) has announced that its wholly-owned subsidiary, Jiangsu YunGongChang Technology Information Technology Co., Ltd., has won a bid for the second phase of the artificial intelligence industrial base project through the Anhui Provincial Bidding Procurement Association [1] Group 1: Project Details - The procurement contract was signed on January 30, 2026, with Suzhou Huari Network Information Service Co., Ltd., which acts as the buyer [1] - The contract value is set at RMB 520 million [1] Group 2: Strategic Focus - During the procurement and implementation phase, the company will scale the use of domestic AI computing power cards, demonstrating its commitment to meeting project demands and supporting the domestic computing ecosystem [1] - The company plans to continue deep investments and long-term strategic layout in the domestic computing power sector, enhancing and enriching its heterogeneous computing capabilities [1]

Core Viewpoint - Pingao Co., Ltd. (688227.SH) is expected to achieve stable revenue in 2025, with projected operating income between 450 million to 530 million yuan, and a significant reduction in net loss compared to 2024 [1][2] Group 1: Financial Performance - The company anticipates a net profit attributable to shareholders narrowing to a loss of 10.0552 million to 20.0552 million yuan [1] - The net cash flow from operating activities is expected to reach 24 million to 36 million yuan, marking a substantial increase of 136.84 million to 148.84 million yuan, with a growth rate of 121.27% to 131.90% [1] - The improvement in performance is attributed to a dual strategy of cost reduction and enhanced receivables management, leading to a decrease in credit impairment losses [1] Group 2: Strategic Developments - The company is accelerating its strategic layout in the artificial intelligence and domestic computing power ecosystem, with significant progress in cloud operating systems for government sectors, serving 12 provincial-level government cloud platforms with over 60% domestic replacement rate [2] - AI-related businesses have made substantial breakthroughs, with products like AIStack, AIInfra, and the "Pinyuan AI Integrated Machine" contributing to revenue [2] - The Pinyuan AI Integrated Machine (PYD20 series) is the first domestically produced AI inference product with a dual-chip architecture, enhancing model inference efficiency and successfully replacing foreign products in various sectors [2] Group 3: Future Outlook - The company plans to deepen its dual-driven strategy of "vertical field AI + domestic computing power ecosystem," focusing on core markets such as government and public security while accelerating AI integration and ecosystem co-construction [2]

Core Insights - The announcement from Beijing Zhiyuan Huazhang Technology Co., Ltd. highlights a significant increase in user demand for AI models, leading to a temporary strain on computing resources, reflecting a broader trend in the AI industry [1] - The AI chip market in China is projected to exceed 1 trillion yuan by 2028, accounting for approximately 30% of the global market, emphasizing the need for high-quality, domestically controlled AI computing power to seize opportunities in the AI sector [1] Group 1: Current State of Computing Power - There is a pronounced supply-demand imbalance in computing power, particularly in China, which is more evident than in the global context [3] - Foreign companies dominate the global AI computing power market, holding nearly 70% of the market share in China by 2024, creating a substantial self-sufficiency gap for domestic production [4] - The self-sufficiency rate of AI GPUs in China has increased from less than 10% in 2020 to approximately 34% in 2024, with expectations to reach around 82% by 2027 [4] Group 2: Causes of Supply-Demand Imbalance - The supply of computing resources in China faces multiple constraints, including limitations on high-end chip imports and performance gaps between domestic and international GPU products [5] - The fragmentation of computing resources among service providers leads to low utilization rates, exacerbating the supply-demand mismatch [5] - The rapid deployment of AI applications across various industries has resulted in over 13,000 projects and more than 30,000 smart factories, significantly increasing the demand for computing power [5] Group 3: Solutions to Computing Power Challenges - To address the computing power challenges, it is essential to maximize the potential of domestic computing resources and promote their application [6] - The Chinese government has been actively implementing policies to enhance computing power development, including optimizing infrastructure and improving service levels [6] - Collaboration among stakeholders in the computing power ecosystem is crucial for achieving deep integration of models, applications, and computing resources, which will enable China to gain a competitive edge in the global AI landscape [7]

Group 1 - The core strategy of Pingao Co., Ltd. is centered around the "Cloud Data Base Platform," focusing on a dual-driven approach of "vertical field artificial intelligence + domestic computing power ecosystem" [1] - The company is gradually building a diversified business system that covers full-stack cloud services, cloud-edge-end collaborative bases, and a domestic computing power ecosystem [1] Group 2 - Shenzhen Jiangyuan Technology Co., Ltd. specializes in the production of high-performance training and inference chips domestically, focusing on heterogeneous computing scenarios [1] - The company is involved in the research, development, and sales of general-purpose computing chips and supporting solutions [1]

Group 1 - The company, Donghua Software, confirmed it is not involved in the development of the MUSA system by Moore Threads, but it is an important ecological partner in the project [2] - The collaboration focuses on advancing the domestic computing power ecosystem [2] - The company emphasized that significant business information should be referenced from official announcements [2]

Core Insights - Paradigm Intelligence recently announced that its "ModelHub XC" has completed the adaptation certification of 108 mainstream AI models on MoEr Thread's GPU, covering various task types such as text generation, visual understanding, and multimodal Q&A, with plans to expand to a thousand models in the next six months [1][3] - MoEr Thread, a domestic GPU company that went public on the Sci-Tech Innovation Board this year, demonstrated significant advantages in quantized models during the adaptation process, effectively reducing model memory usage and improving inference speed [1] - The official launch of MoEr Thread's IPO on November 24 set a new high for A-share IPO prices since 2025, with a price of 114.28 yuan per share [1] - The efficient and stable operation of models on domestic chips is crucial for the maturity of the computing power ecosystem, and Paradigm Intelligence is addressing this by leveraging its self-developed EngineX engine technology to enhance model compatibility and operational efficiency on domestic chips [1] Model Adaptation and EngineX Technology - The "ModelHub XC" has successfully validated model series including Mata, Qianwen, Deepseek, and others on MoEr Thread's GPU, with the EngineX engine enabling "engine-driven, multi-model plug-and-play" capabilities [3] - EngineX serves as the underlying support system, effectively addressing the bottlenecks in model compatibility and scalability on domestic chips [3] Overview of ModelHub XC - ModelHub XC is an AI model and tool platform aimed at the domestic computing power ecosystem, combining community and service functions to promote AI innovation and implementation on domestic hardware platforms, providing comprehensive solutions from model training to inference and deployment [6]

Core Viewpoint - The semiconductor sector in China is experiencing fluctuations, with the STAR Market Chip Index showing a slight decline, while certain stocks are performing well, indicating mixed market sentiment in the industry [1] Industry Summary - As of December 8, U.S. chip stocks saw gains post-market, with NVIDIA rising nearly 3%, reflecting positive trends in the global semiconductor market [1] - Dongguan Securities highlights that artificial intelligence remains a key innovation driver in the tech sector, with various segments such as computing power, storage, equipment, and advanced packaging expected to benefit [1] - Domestic AI chip companies have rapidly developed and achieved significant progress in localization, with firms like Moore and Muxi accelerating their capital market strategies [1] - Major internet companies, including Tencent, are actively adapting to domestic computing power chips, which is expected to expedite the formation of a domestic computing power ecosystem [1] Company Summary - As of November 28, the top ten weighted stocks in the STAR Market Chip Index include Haiguang Information, Cambrian, SMIC, and others, collectively accounting for 59.66% of the index [1] - The STAR Chip ETF (588200) serves as a convenient tool for investors looking to gain exposure to the STAR Market chip sector [1] - Investors without stock accounts can access investment opportunities in domestic chips through the STAR Chip ETF linked fund (017470) [1]