Investment Rating - The report maintains a positive outlook on AI applications, Huawei chain, and financial technology [2][5][12] Core Viewpoints - The report highlights the investment hotspots in July, focusing on overseas computing power, AI applications, and military industry, with a bullish sentiment in the market [2][5][12] - The report emphasizes the importance of technological innovation as a growth driver, particularly in AI applications, Huawei chain, and financial technology [2][5][12] Summary by Sections July Investment Hotspots Review - The report identifies overseas computing power and AI applications as key investment themes, with significant gains observed in these sectors [23][34] - The report notes that the computing sector has shown strong performance, with the Shenwan Computer Index rising by 3.86% in July [12][14] August Investment Direction - The report suggests focusing on three main directions for August: AI applications, Huawei chain, and financial technology [2][5][12] - AI applications are supported by government policies and advancements in large model iterations, with notable companies like TempusAI expected to report strong earnings [2][5][12] - The Huawei chain is anticipated to benefit from the upcoming Huawei Connect conference, with historical data indicating potential excess returns around such events [2][5][12] - Financial technology is highlighted as a leading sector, with a significant increase in new A-share accounts and trading volume, indicating a bullish market sentiment [2][5][12] Key Companies and Performance - The report lists several companies with notable performance in July, including Yao Cai Securities, Da Zhi Hui, and OSL Group, with gains exceeding 30% [13][34] - AI application companies such as Dingjie Zhizhi and Fanwei also showed strong performance, with significant month-over-month increases [34][35] Policy Support for AI Development - The report discusses recent government initiatives aimed at supporting AI development, including the issuance of AI vouchers and funding for key projects [35][36] - Shanghai's measures to expand AI applications include financial incentives for companies involved in AI technology and infrastructure [36][37]

Core Insights - The Huawei Cloud City Summit 2025 was successfully held in Wuxi, focusing on AI applications, large models, and embodied intelligence technology [1] - The establishment of the Artificial Intelligence Innovation Center in Wuxi aims to integrate technology research and development, industry empowerment, and ecosystem cultivation [1] - Huawei Cloud emphasizes the importance of AI as a transformative technology across various industries, aiming to reshape sectors through secure and innovative cloud services [1] Group 1 - Huawei Cloud introduced the CloudMatrix 384 super-node architecture to meet the massive computing power demands of the AI era, transforming resource supply models from server-level to matrix-level [3] - The CloudMatrix 384 super-node features high density, high speed, and high efficiency, achieving comprehensive leadership in computing power, interconnect bandwidth, and memory bandwidth through architectural innovation [3] Group 2 - Huawei Cloud aims to empower partners and clients by leveraging the Pangu large model, focusing on creating value through key enterprise scenarios [4] - China holds a leading position in the industrial sector, with the manufacturing industry being the largest globally for 14 consecutive years, providing a strong foundation for innovation and intelligent upgrades [4] - Huawei Cloud's strategy includes a three-pronged approach of foundational models, toolchains, and industry scenarios to provide ready-to-use large model capabilities across over 30 industries [4] Group 3 - The Artificial Intelligence Innovation Center in Wuxi aims to support industries such as integrated circuits, high-end equipment, new energy, biomedicine, new materials, automotive parts, information technology, and aerospace in accelerating intelligent upgrades [5] - The center seeks to gather internal best practices and external ecosystem experiences on a cloud platform to empower partners and clients [5]

Core Viewpoint - Huawei's Ascend CloudMatrix 384 super node has demonstrated performance that surpasses NVIDIA's products in certain aspects, indicating a significant advancement in domestic AI chip capabilities [1][11][13]. Group 1: Huawei's Ascend Chip Overview - Ascend is a dedicated AI processing chip (NPU) designed specifically for AI tasks, with the Ascend 910 being its main product [3][6]. - Previously, Ascend chips were used as backup options due to the unavailability of high-end NVIDIA and AMD chips, but they have now emerged as leaders in the domestic chip market [3][6]. - The Ascend chips have primarily been utilized in AI inference tasks, with limited use in model training due to performance and ecosystem limitations [4][6]. Group 2: Performance and Capabilities - In 2024 and 2025, Huawei transformed Ascend from a backup option to a primary player capable of training large models, achieving significant results documented in research papers [5][6]. - Ascend has successfully trained models with 135 billion parameters using 8192 chips and 718 billion parameters using over 6000 chips, showcasing the ability to train large-scale models with domestic chips [6][10]. - Key performance indicators such as MFU (Modeling Function Utilization) reached over 50% for the dense model and 41% for the MoE model, indicating high efficiency in resource utilization [9][10]. Group 3: Competitive Comparison with NVIDIA - In direct comparisons, Ascend's 384 super node demonstrated comparable performance to NVIDIA's H100 and H800 in real-world applications, achieving the best utilization rates [11][12]. - Although a single Ascend chip's performance is only one-third of NVIDIA's Blackwell, the overall system performance of the 384 super node exceeds NVIDIA's GB200 due to the higher number of chips used [13][21]. - This indicates that Ascend is not just a replacement but has the potential to lead in certain performance metrics [13]. Group 4: Technological Innovations - The CloudMatrix 384 super node consists of 384 Ascend 910 chips and 192 Kunpeng CPUs, interconnected using advanced optical communication technology, which enhances data transmission efficiency [16][30]. - Huawei's approach focuses on a system-level engineering breakthrough rather than relying on single-chip performance, utilizing a combination of communication, optical, thermal, and software innovations [21][22]. - The architecture allows for high-speed, peer-to-peer communication among chips, significantly improving data transfer rates compared to traditional copper connections used by competitors [28][30]. Group 5: Market Position and Future Outlook - Despite still trailing behind NVIDIA in chip technology and software ecosystem, Huawei's Ascend has gained traction in the Chinese market, especially as companies adapt to domestic chips due to restrictions on NVIDIA products [36][38]. - The domestic semiconductor industry is evolving under pressure, with Huawei's strategy representing a unique "technology curve" that prioritizes system optimization over individual chip performance [38][39]. - The advancements made by Ascend may signify the beginning of a significant shift in the AI computing landscape, positioning domestic capabilities for a potential resurgence in the global market [40].

Core Viewpoint - Huawei's Ascend is positioning itself as a competitive alternative to NVIDIA by leveraging a system-level engineering breakthrough rather than focusing solely on individual chip performance [5][12]. Technical Breakdown - The "CloudMatrix 384 Super Node" consists of 384 Ascend 910 chips and 192 Kunpeng CPUs, interconnected to function as a single efficient unit [2][4]. - Unlike NVIDIA's reliance on copper cables for inter-chip communication, Huawei has developed a fully peer-to-peer interconnect bus using optical cables, significantly enhancing data transmission efficiency [8][10]. - Huawei's approach includes advanced mathematical algorithms for scheduling, deep hardware-software collaboration, and effective thermal management, contributing to the overall performance of the Ascend system [11]. Competitive Landscape - Huawei acknowledges a gap in chip technology, as NVIDIA has advanced to 3nm processes, while Huawei's chips are still a generation behind [14]. - The software ecosystem remains a challenge, with Huawei's CANN still needing to catch up to NVIDIA's established CUDA framework [15]. - Despite these challenges, Huawei is gaining traction in the Chinese market, especially as companies adapt to domestic chips due to restrictions on NVIDIA's products [16]. Industry Dynamics - The domestic AI chip market is diversifying into three main factions: tech giants like Huawei, pure chip manufacturers, and niche players focusing on specific applications [16]. - The pressure from U.S. sanctions has inadvertently spurred growth in China's semiconductor industry, allowing companies like Huawei to innovate and optimize system performance rather than solely competing on chip specifications [16][17].

Core Viewpoint - Huawei Cloud Stack aims to provide a hybrid cloud solution that better understands the needs of government and enterprise users in the era of intelligence, focusing on AI integration and data utilization [1][3]. Group 1: Huawei Cloud Stack Features - Huawei Cloud Stack will become the first hybrid cloud to adapt to CloudMatrix 384 super nodes, enabling enterprise customers to have their own cloud super nodes locally, enhancing AI computing power for intelligent transitions [3]. - Currently, Huawei Cloud Stack offers over 120 cloud services and more than 50 scenario-based solutions, maintaining the leading market share in the hybrid cloud sector across government, finance, and manufacturing for several consecutive years [3][4]. Group 2: User Segmentation and Solutions - Huawei Cloud Stack recognizes that government and enterprise users are not a monolithic group but can be categorized into four distinct roles: data center engineers, data engineers, AI algorithm model application engineers, and application development engineers [3][4]. - The platform supports users throughout the entire cloud lifecycle, from building to managing cloud resources, enabling efficient resource allocation, data governance, model training, and application development [4]. Group 3: Case Studies - In finance, Huawei Cloud Stack helped a state-owned bank establish a unified computing power platform, allowing data center engineers to deploy 106 DeepSeek R1 instances in just two days, improving efficiency by 70% compared to traditional bare-metal deployments [4][5]. - In manufacturing, Huawei Cloud collaborated with XCMG to create a robust big data platform, enhancing data analysis efficiency and enabling value extraction from operational data of construction machinery [4][5]. - In the steel industry, Xianggang utilized Huawei Cloud Stack to develop a one-stop AI development platform, achieving quality improvement and cost reduction through the deployment of a steel model across over 30 scenarios [5]. - In the energy sector, CNOOC implemented CodeArts to develop a digital platform, reducing development time by 30% and streamlining the deployment of intelligent oilfield management systems from one week to one day [5].

Core Viewpoint - The article discusses the advancements in AI computing clusters, emphasizing their critical role in enhancing the capabilities of AI models through innovative engineering solutions and fault tolerance mechanisms [1]. Group 1: Supernode High Availability - AI training and inference require continuous operation, with each computer in the cluster having a backup to ensure seamless task execution during failures [1]. - Huawei's fault tolerance solutions include system-level, business-level, and operational-level strategies to manage faults gracefully [1]. Group 2: Cluster Linearity - The ideal scenario for computing clusters is linear scalability, where the performance increases proportionally with the number of computers [1]. - Huawei employs advanced task allocation algorithms and technologies to achieve high linearity in model training, with results showing linearity rates of 96% for various configurations [1]. Group 3: Rapid Recovery in Large-Scale Training - The system can automatically save training progress, allowing for quick recovery from failures without starting over [1]. - Innovations include process-level rescheduling and online recovery techniques that significantly reduce recovery times to under 3 minutes [1]. Group 4: Large-Scale MoE Model Inference Recovery - The article outlines a three-tier fault tolerance strategy for large-scale MoE model inference, minimizing user impact during hardware failures [1]. - Techniques such as rapid instance restart and token-level retries have been validated to reduce recovery times significantly [1]. Group 5: Fault Management and Diagnostic Awareness - A real-time monitoring system continuously tracks the health of each computer in the cluster, enabling quick fault detection and diagnosis [1]. - Huawei's comprehensive fault management solutions enhance reliability through advanced diagnostic capabilities and proactive maintenance strategies [1]. Group 6: Simulation Modeling - The article introduces a Markov modeling simulation platform that allows for pre-testing of AI models in a virtual environment, identifying potential bottlenecks before real-world deployment [1]. - This approach optimizes resource allocation and enhances the overall efficiency of the computing cluster [1]. Group 7: Framework Migration - Huawei's MindSpore framework supports seamless integration with mainstream ecosystems, facilitating the deployment of large models and improving inference performance [1]. - The framework includes tools for adapting third-party frameworks, ensuring compatibility and efficiency in AI model training and inference [1].

Core Viewpoint - The article discusses the rapid development and deployment of the DeepSeek AI model and the CloudMatrix 384 super node by Huawei Cloud, highlighting the competitive landscape and technological advancements in China's AI infrastructure [5][8][76]. Group 1: DeepSeek and Competitive Landscape - The launch of DeepSeek R1 on January 15, 2025, created a significant impact globally, leading to intense competition among AI teams worldwide [5][6]. - Just two weeks later, Huawei Cloud partnered with Silicon-based Flow to launch DeepSeek R1/V3, marking a major milestone in domestic AI deployment [8][9]. - The collaboration aimed to leverage domestic computing power to address deployment challenges, resulting in a 40-fold increase in website traffic for Silicon-based Flow in February [14][15]. Group 2: CloudMatrix 384 Super Node - Huawei Cloud introduced the CloudMatrix 384 super node, which integrates 384 Ascend cards, significantly surpassing NVIDIA's NVL72 super node, which integrates only 72 GPUs [24][21]. - The CloudMatrix 384 is designed to meet the high demands of AI applications, requiring substantial power and advanced cooling systems to manage increased heat density [45][46]. - Despite initial doubts about the feasibility of using optical modules for communication, Huawei Cloud successfully addressed challenges related to communication quality and stability, leading to a successful deployment of the super node [49][56]. Group 3: Performance Improvements - By April 10, 2025, the performance of the DeepSeek service on the CloudMatrix 384 improved dramatically, achieving a throughput of 1920 Tokens/second, which is six times better than earlier performance [69][71]. - The enhancements included significant improvements in various performance metrics, such as training and inference efficiency, showcasing Huawei Cloud's advancements in AI technology [71][75]. - The successful deployment of the CloudMatrix 384 super node represents a breakthrough in China's AI capabilities, establishing a robust domestic computing power infrastructure [76][78].

Core Viewpoint - Reinforcement Learning (RL) post-training has become a crucial method for breaking through the performance ceiling of large language models (LLMs), with Huawei introducing two key technologies to enhance efficiency and resource utilization in this process [1][2][26]. Group 1: RL Post-Training Technologies - RL post-training now consumes 20% of the total computational power in the training process, projected to rise to 50%, significantly impacting model performance and costs [1]. - Huawei's "RL Fusion" technology allows a single card to handle both training and inference tasks simultaneously, doubling resource utilization and throughput [4][5]. - The "StaleSync" mechanism breaks the synchronization limitations, achieving over 90% efficiency in cluster expansion and a 50% increase in training throughput [2][10]. Group 2: Challenges in RL Post-Training - The traditional On-Policy algorithm requires alternating between training and inference tasks, leading to significant resource idling, especially in large-scale clusters [3]. - The complexity of task scheduling has increased exponentially with the popularity of Mixture of Experts (MoE) models, complicating resource utilization [4]. Group 3: Performance Improvements - RL Fusion technology enables dynamic switching between training and inference modes, optimizing memory usage and enhancing efficiency [5][8]. - The combination of RL Fusion and StaleSync technologies has led to a 78.5% increase in throughput for single supernodes, with overall performance improvements of 1.5 times [22][24]. - StaleSync allows for linear scalability in cluster expansion, with throughput increasing from 35k tokens/s to 127k tokens/s as the cluster size grows from one to four supernodes, achieving a linearity of 91% [24]. Group 4: Future Implications - The advancements in RL post-training technologies position Huawei's CloudMatrix 384 supernodes as a "super accelerator" for training large models, enhancing speed and efficiency significantly [2][26]. - The innovations in resource utilization and task parallelism are expected to drive the next generation of AI efficiency, marking a pivotal moment in the evolution of large model training [26].

Core Viewpoint - Reinforcement Learning (RL) post-training has become a crucial path for breaking through the performance ceiling of large language models (LLMs), with Huawei introducing two key technologies to enhance efficiency and resource utilization in this process [2][3][56]. Group 1: RL Post-Training Challenges - RL post-training currently consumes 20% of the total computational power in the training process, projected to rise to 50%, significantly impacting model performance and costs [3]. - Traditional RL post-training suffers from low resource utilization due to the alternating execution of training and inference tasks, leading to substantial computational waste [11][13]. - The complexity of task scheduling in large-scale clusters has increased due to the popularity of Mixture of Experts (MoE) models, making efficient collaboration challenging [15][16]. Group 2: Huawei's Innovations - Huawei's "RL Fusion" technology allows a single card to handle both training and inference tasks simultaneously, effectively doubling resource utilization and throughput [5][18]. - The "StaleSync" mechanism enables a quasi-asynchronous approach, allowing different RL tasks to execute in parallel within a defined "staleness threshold," improving horizontal scaling efficiency to over 90% [29][32]. - The combination of RL Fusion and StaleSync technologies significantly enhances the efficiency of RL post-training, achieving a throughput increase of 1.5 times [52][56]. Group 3: Performance Metrics - The implementation of RL Fusion can lead to a throughput increase from 14.0k tokens/sec to 35.0k tokens/sec when combined with StaleSync, representing a 150% improvement compared to baseline configurations [54]. - In a multi-node setup, StaleSync allows for linear scaling efficiency, with throughput increasing from 35k tokens/sec to 127k tokens/sec as the number of nodes increases from 1 to 4, achieving a linearity of 91% [55].

Core Viewpoint - Huawei has achieved significant advancements in training large models through its "Ascend + Pangu Ultra MoE" system, demonstrating a fully domestic and GPU-free training process that enhances computational efficiency and model performance [3][4][38]. Group 1: Technical Innovations - Huawei's training system has achieved a model training efficiency with a utilization rate (MFU) of 41% during the pre-training phase using the Ascend Atlas 800T A2 cluster [4][38]. - The Pangu Ultra MoE model consists of 718 billion parameters, featuring a unique architecture with 61 layers, including 58 MoE layers, and is designed for high performance and scalability [38][39]. - The system supports a high throughput of 35K Tokens/s during the reinforcement learning (RL) post-training phase, showcasing its capability to process complex tasks rapidly [39]. Group 2: Challenges Addressed - The report identifies six key challenges in the current MoE pre-training and RL post-training processes, including difficulties in parallel strategy configuration, communication bottlenecks, and uneven system load distribution [7][10][12][13]. - Huawei has developed a comprehensive end-to-end solution to address these challenges, focusing on optimizing training cluster utilization and enhancing communication efficiency [14][16][25]. Group 3: Specific Solutions - The first strategy involves improving training cluster utilization through intelligent parallel strategy selection and global dynamic load balancing, significantly enhancing overall training efficiency [16][23]. - The second strategy focuses on releasing computational power at the single-node level by optimizing training operators and enhancing memory management, achieving a twofold increase in micro-batch size [26][30]. - The third strategy introduces high-performance scalable RL post-training technologies, allowing for flexible deployment modes and doubling the utilization rate of RL post-training clusters [33][34].