Investment Rating - The report does not provide a specific investment rating for the industry. Core Insights - The development of large language models (LLMs) has evolved significantly, with key milestones including the introduction of the Transformer architecture by Google in 2018 and the release of models like GPT-3 and GPT-4, which have billions of parameters and demonstrate emergent capabilities [4][28][37]. - LLMs are transforming various sectors, including natural language processing, information retrieval, computer vision, and the development of AI agents, indicating their potential as foundational models for diverse applications [7][12]. - The emergence of capabilities in LLMs allows them to perform complex tasks with minimal data, showcasing their efficiency and adaptability in various contexts [11][12]. Summary by Sections Language Model Development - The history of language models dates back to the 1990s, with significant advancements in deep learning integration and the introduction of transformer architectures [4][32]. - Notable models include GPT-3 with 175 billion parameters and GPT-4, which further enhances capabilities and introduces multimodal understanding [28][37]. Impact on Technology and Business - LLMs enhance natural language processing tasks such as text generation, translation, and question answering, while also improving information retrieval systems [7][12]. - The models support various applications, including digital assistants and emotional analysis, indicating their broad utility in commercial settings [7][12]. Emergent Capabilities - LLMs exhibit emergent abilities, allowing them to tackle new tasks with limited examples, which reduces the need for extensive retraining [11][12]. - The models leverage vast amounts of unlabelled data for training, enabling them to generalize across multiple downstream tasks effectively [11][12]. Model Training and Architecture - The training of LLMs involves pre-training on large datasets followed by fine-tuning for specific tasks, which enhances their performance across various applications [12][28]. - The architecture of these models, particularly the use of transformers, allows for efficient processing of language and context, leading to improved understanding and generation capabilities [4][32]. Future Directions - The report highlights ongoing research and development in LLMs, with a focus on improving their efficiency, ethical considerations, and addressing challenges such as data privacy and bias [12][28]. - The industry is witnessing a trend towards more accessible and versatile models, with companies like OpenAI, Google, and Baidu leading the charge in developing advanced LLMs [37][47].

Core Insights - The article discusses Google's TPU v7 interconnect architecture, focusing on the ratio of TPU to copper cables and optical modules, highlighting the technical aspects of the TPU design and its cooling solutions [1][6][7]. TPU Rack Interconnect Architecture - One of the notable features of TPU is its ability to achieve large-scale world size expansion through the ICI protocol, with a TPU Pod capable of accommodating up to 9216 Ironwood TPUs [2]. - Each TPU rack consists of 16 TPU trays and a varying number of host CPU trays, along with a top-of-rack switch and power units [2]. - The TPU tray contains a TPU board with four TPU chips, each equipped with multiple interfaces for interconnectivity [2]. Cooling Solutions - Google has adopted liquid cooling for TPU racks since the TPU v3 era, with a 1:1 ratio of TPU trays to host CPU trays in liquid-cooled racks, compared to a 2:1 ratio in air-cooled racks [6]. - The market anticipates that 2024 will be the "year of liquid cooling," as more ASIC servers begin to adopt this technology, indicating significant market growth potential [6]. Market Projections - In 2026, Google is expected to ship 2.5 million TPU v7 units, leading to a liquid cooling market space of approximately $2.8 to $3.2 billion [7]. - By 2027, shipments are projected to exceed 5 million units, with the value of liquid cooling per rack potentially increasing to $90,000 to $100,000, resulting in a market space of $7 to $8 billion [7]. Interconnect Design - The TPU v7 utilizes a 3D torus topology for interconnectivity, where each TPU connects to six neighboring nodes across three dimensions [8]. - Internal connections within the TPU tray use copper cables, while external connections utilize optical modules and OCS for inter-unit communication [9][12]. Optical Connectivity and Market Demand - A TPU Pod with 9216 TPUs will require approximately 11,520 copper cables and 13,824 optical modules, indicating a significant demand for optical components in the market [16]. - Google is projected to need around 15,000 OCS switches by 2026, with a market space for OCS estimated at $2.2 billion based on a price of $150,000 per switch [17][18].

Investment Rating - The report maintains an "Accumulate-A" rating for the company [2] Core Views - The company reported a revenue of 430 million yuan in the first half of 2025, representing a year-on-year growth of 5.9%, but the net profit attributable to the parent company decreased by 48.2% to 30 million yuan [4] - The decline in performance is attributed to a decrease in telecom transmission demand and insufficient release of DCI capacity [5] - The company is expected to accelerate the release of DCI capacity in the second half of the year, with projections for net profits of 90 million, 290 million, and 590 million yuan for 2025, 2026, and 2027 respectively [8] Financial Performance - The company achieved a gross margin of 26.3% in the first half of 2025, down 5.2 percentage points year-on-year [4] - The revenue from the transmission product line was 330 million yuan, a year-on-year decrease of 7.9%, while the access and data product lines saw a significant increase of 104.7% to 100 million yuan [5] - The projected revenue for 2025 is 1.177 billion yuan, with a year-on-year growth of 39.9% [10] Market Trends - The global DCI market is expected to reach over 40 billion USD in 2025, growing by 14.3% year-on-year, driven by increased demand for data center connectivity and distributed training [6] - The company has received sample orders for its silicon-based OCS optical switch, indicating potential for future mass production [7] Profitability Forecast - The report forecasts a decline in net profit for 2025, with a projected net profit margin of 7.6% [10] - The company's earnings per share (EPS) for 2025 is estimated at 0.57 yuan, with a price-to-earnings (P/E) ratio of 247.3 [10]

Core Viewpoint - The article discusses the architecture and data flow of Huawei's Cloud Matrix 384, emphasizing the integration of optical and electrical interconnections in its network design [2][3][9]. Group 1: Data Transmission Layers - The Cloud Matrix 384 includes three main data transmission layers: UB Plane, RDMA Plane, and VPC Plane, each serving distinct roles in data processing and communication [5][7]. - The UB Plane connects all NPU and CPU with a non-blocking full-mesh topology, providing a unidirectional bandwidth of 392GB/s per Ascend 910C [7]. - The RDMA Plane facilitates horizontal scaling communication between supernodes using RoCE protocol, primarily connecting NPUs for high-speed KV Cache transfer [7]. - The VPC Plane connects supernodes to broader data center networks, managing tasks such as storage access and external service communication [7]. Group 2: Optical and Electrical Interconnections - Although the Cloud Matrix 384 is often referred to as a purely optical interconnection system, it also utilizes electrical interconnections for short distances to reduce costs and power consumption [9]. - The article highlights the necessity of both optical and electrical connections in achieving efficient data flow within the system [9]. Group 3: Scale-Up and Scale-Out Calculations - For Scale-Up, each server's UB Switch chip corresponds to a bandwidth of 448GBps, requiring 56 400G optical modules or 28 800G dual-channel optical modules per server [12]. - The ratio of NPUs to 400G optical modules in Scale-Up is 1:14, and to 800G modules is 1:7 [12]. - For Scale-Out, a Cloud Matrix node consists of 12 Compute cabinets, and the optical module demand ratio is approximately 1:4 for NPUs to 400G optical modules [14].

Core Viewpoint - The article discusses the competition between InfiniBand and Ethernet in AI networking, highlighting the advantages of Ethernet in terms of cost, scalability, and compatibility with existing infrastructure [6][8][22]. Group 1: AI Networking Overview - AI networks are primarily based on InfiniBand due to NVIDIA's dominance in the AI server market, but Ethernet is gaining traction due to its cost-effectiveness and established deployment in large-scale data centers [8][20]. - The establishment of the "Ultra Ethernet Consortium" (UEC) aims to enhance Ethernet's capabilities for high-performance computing and AI, directly competing with InfiniBand [8][9]. Group 2: Deployment Considerations - Teams face four key questions when deploying AI networks: whether to use existing TCP/IP networks or build dedicated high-performance networks, whether to choose InfiniBand or Ethernet-based RoCE, how to manage and maintain the network, and whether it can support multi-tenant isolation [9][10]. - The increasing size of AI models, often reaching hundreds of billions of parameters, necessitates distributed training, which relies heavily on network performance for communication efficiency [10][20]. Group 3: Technical Comparison - InfiniBand offers advantages in bandwidth and latency, with capabilities such as high-speed data transfer and low end-to-end communication delays, making it suitable for high-performance computing [20][21]. - Ethernet, particularly RoCE v2, provides flexibility and cost advantages, allowing for the integration of traditional Ethernet services while supporting high-performance RDMA [18][22]. Group 4: Future Trends - In AI inference scenarios, Ethernet is expected to demonstrate greater applicability and advantages due to its compatibility with existing infrastructure and cost-effectiveness, leading to more high-performance clusters being deployed on Ethernet [22][23].

Core Insights - The advancement of artificial intelligence (AI) relies on the exponential growth of AI supercomputers, with training compute power increasing by 4.1 times annually since 2010, leading to breakthroughs in various AI applications [1][13] - The performance of leading AI supercomputers doubles approximately every nine months, driven by a 1.6 times annual increase in the number of chips and their performance [2][3] - By 2025, the most powerful AI supercomputer, xAI's Colossus, is estimated to have a hardware cost of $7 billion and a power demand of around 300 megawatts, equivalent to the electricity consumption of 250,000 households [3][41] Group 1: AI Supercomputer Performance and Growth - The performance of leading AI supercomputers is projected to grow at an annual rate of 2.5 times, with private sector systems growing even faster at 3.1 times [21][29] - The number of AI chips in top supercomputers is expected to increase from over 10,000 in 2019 to over 200,000 by 2024, exemplified by xAI's Colossus [2][24] - The energy efficiency of AI supercomputers is improving, with a yearly increase of 1.34 times, primarily due to the adoption of more energy-efficient chips [45][49] Group 2: Hardware Costs and Power Demand - The hardware costs of leading AI supercomputers are projected to double annually, reaching approximately $2 billion by 2030 [50][73] - Power demand for these supercomputers is expected to grow at a rate of 2.0 times per year, potentially reaching 9 gigawatts by 2030, which poses significant challenges for infrastructure [41][75] - The rapid increase in power demand may lead companies to adopt distributed training methods to manage workloads across multiple locations [76][77] Group 3: Market Dynamics and Geopolitical Implications - The private sector's share of AI supercomputer performance has surged from under 40% in 2019 to about 80% by 2025, while the public sector's share has dropped below 20% [8][56] - The United States dominates the global AI supercomputer landscape, accounting for approximately 75% of total performance, followed by China at 15% [10][59] - The shift from public to private ownership of AI supercomputers reflects the growing economic importance of AI and the increasing investment in AI infrastructure [54][68]