中信证券股份有限公司 关于深圳云天励飞技术股份有限公司 首次公开发行部分限售股上市流通 暨控股股东、实际控制人及其一致行动人自愿承诺不减持的核查意见 中信证券股份有限公司（以下简称"保荐人"）为深圳云天励飞技术股份有 限公司（以下简称"公司"、"云天励飞"）首次公开发行股票并在科创板上市项 目的保荐人。根据《中华人民共和国公司法》《中华人民共和国证券法》《证券发 行上市保荐业务管理办法》《上海证券交易所科创板上市公司自律监管指引第 1 号——规范运作》《上海证券交易所科创板股票上市规则》以及《科创板上市公 司持续监管办法（试行）》等有关规定，对云天励飞首次公开发行部分限售股上 市流通暨控股股东、实际控制人及其一致行动人自愿承诺不减持事项进行了核查， 核查情况如下： 一、本次上市流通的限售股类型 根据中国证券监督管理委员会于 2023 年 1 月 4 日出具的《关于同意深圳云 天励飞技术股份有限公司首次公开发行股票注册的批复》（证监许可〔2023〕13 号），公司首次向社会公开发行人民币普通股（A 股）股票 8,878.3430 万股，并 于 2023 年 4 月 4 日在上海证券交易所科创板挂牌上市。公司发行 ...

证券代码：688343 证券简称：云天励飞 公告编号：2026-006 深圳云天励飞技术股份有限公司 首次公开发行部分限售股上市流通暨控股股东、实际控 制人及其一致行动人自愿承诺不减持的公告 本公司董事会及全体董事保证本公告内容不存在任何虚假记载、误导性陈述 或者重大遗漏，并对其内容的真实性、准确性和完整性依法承担法律责任。 重要内容提示： 本次股票上市类型为首发限售股份；股票认购方式为网下，上市股数为 94,996,440股。 本次股票上市流通总数为94,996,440股。 本次股票上市流通日期为2026 年 4 月 7 日。（因 2026 年 4 月 4 日为非交易 日，上市流通日顺延） 深圳云天励飞技术股份有限公司（以下简称"公司"或"云天励飞"或"发 行人"）控股股东、实际控制人及其一致行动人陈宁先生及珠海明德致远投资有限 公司自愿承诺：自 2026 年 4 月 7 日（本次限售股上市流通日）起 12 个月内，不 以任何形式减持其直接持有的公司首次公开发行前股份 89,755,780 股，占本次上 市流通总数的 94.48%，占公司总股本的 24.96%。 一、本次上市流通的限售股类型 三、本次上市流 ...

Investment Rating - The report assigns a "Buy" rating for the industry, predicting that the stock price will outperform the Shanghai Composite Index by 15% or more within the next six months [49]. Core Insights - OpenClaw is an open-source autonomous AI virtual assistant software project designed to perform complex tasks autonomously based on user instructions, marking a significant advancement from traditional chatbots [4][15]. - The demand for AI agents is expected to accelerate, with the number of agents in Chinese enterprises projected to exceed 350 million by 2031, achieving a compound annual growth rate of over 135% [5][24]. - The release of GPT-5.4 enhances the capabilities of AI agents, allowing them to perform complex workflows across software applications, which is crucial for the commercial viability of OpenClaw [5][28][31]. - OpenClaw's rise is anticipated to have three major impacts on the industry: it will likely become one of the first application scenarios for AI agents, increase token consumption significantly, and boost the demand for domestic large models [6][34]. Summary by Sections 01 OpenClaw: A New Era of AI Employees - OpenClaw is an open-source AI agent project developed by Peter Steinberger, initially released as Clawdbot in late 2025 and later renamed [11]. - It is designed to autonomously execute tasks for users, offering capabilities such as file operations, process orchestration, and multi-platform interactions [15][19]. 02 Demand Side: Expansion of Agent Demand - AI agents are seen as a feasible direction for application deployment, with their proactive capabilities distinguishing them from traditional AI assistants [21]. - The growth in agent numbers and token consumption in China is expected to accelerate, driven by advancements in local models and supportive industrial policies [24]. 03 Supply Side: Continuous Iteration of Large Models and Strengthening of Underlying Technologies - The release of GPT-5.4 marks a significant advancement in AI capabilities, particularly in enhancing agent functionalities [28]. - The ongoing improvements in large model technologies are crucial for the commercial success of AI agents like OpenClaw [31]. 04 Impact of OpenClaw's Emergence on the Industry - OpenClaw is gaining rapid popularity, surpassing many established projects on GitHub, indicating its potential as a leading application scenario for AI agents [34]. - The increase in token consumption associated with OpenClaw is expected to drive demand for computing power and cloud services [40]. - The demand for domestic large models is anticipated to rise, with OpenClaw primarily utilizing Chinese models, which may accelerate their international expansion [41]. 05 Investment Recommendations - Beneficial stocks include companies involved in large models and applications, as well as those providing computing infrastructure, indicating a broad investment opportunity in the AI sector [46].

Core Viewpoint - The article discusses the establishment of an AI inference cluster in Zhanjiang by Yuntian Lifei, which aims to enhance AI capabilities in various applications through the use of self-developed domestic AI inference acceleration cards, with a project budget of 420 million yuan [1]. Group 1: AI Inference Shift - AI computing power is transitioning from a "training-first" approach to a "inference-first" model, with inference computing becoming crucial for AI applications [2]. - Gartner predicts that by 2026, approximately 55% of AI-specific cloud infrastructure spending will be allocated to inference workloads [2]. - The Zhanjiang cluster is designed specifically for inference tasks, supporting various industry applications and facilitating the AI transformation of traditional industries [2]. Group 2: Cluster Architecture for Inference Era - The architecture of the inference cluster is designed to meet high concurrency, high throughput, and low latency requirements, utilizing a "Prefill-Decode separation" approach for resource optimization [4]. - The Prefill phase focuses on understanding long contexts, while the Decode phase generates tokens, necessitating efficient resource allocation between the two stages [4]. - Future performance bottlenecks in inference systems are expected to arise from data access efficiency rather than just computational power, highlighting the importance of collaborative design among computing, storage, and networking [4]. Group 3: Self-Developed Chip for Cost-Effective Inference - The AI inference cluster will be built in three phases, all utilizing Yuntian Lifei's self-developed domestic AI inference acceleration cards, with the first phase deploying the X6000 inference acceleration card [7]. - The company plans to release three generations of AI inference chips over the next three years, focusing on optimizing performance for long context scenarios and low-latency decoding [7]. - The long-term goal includes achieving a cost of "one cent per hundred billion tokens" through continuous optimization of chips and systems [7]. Group 4: Industry Implications - The Zhanjiang project represents a shift in AI computing strategy from merely increasing GPU scale to focusing on cost efficiency and stable large-scale inference capabilities [8]. - The establishment of a thousand-card inference cluster not only meets current AI application demands but also serves as a technical deployment platform for larger-scale computing systems [8]. - The collaboration between domestic models and chips is expected to drive the AI infrastructure from technical exploration to large-scale application, opening new avenues for the development of the AI industry [9].

Core Viewpoint - The article discusses the establishment of an AI inference cluster by Yuntian Lifei in Zhanjiang, which aims to create a "national model and national chip" ecosystem, leveraging domestic AI technologies to support various industry applications and enhance local digital transformation [3][14]. Group 1: Project Overview - Yuntian Lifei won a bid for the Zhanjiang AI penetration support project with a contract amount of 420 million yuan, focusing on building a domestic AI inference cluster based on self-developed AI inference acceleration cards [3]. - The cluster will utilize domestic large models like DeepSeek to provide AI capabilities for government and industry applications, aiming to create a model for the "national model and national chip" ecosystem [3][14]. Group 2: Technical Architecture - The AI inference cluster is designed to meet high concurrency, high throughput, and low latency requirements, employing a "Prefill-Decode separation" architecture to optimize resource allocation during different processing stages [6]. - The system architecture prioritizes optimizing the Prefill phase while balancing the Decode phase, ensuring high throughput efficiency even in long-context inference scenarios [7]. Group 3: Chip Development and Cost Efficiency - The AI inference cluster will be built in three phases, all utilizing Yuntian Lifei's self-developed AI inference acceleration cards, with the first phase deploying the X6000 inference acceleration card [10]. - Future plans include launching three generations of AI inference chips over the next three years, focusing on optimizing both Prefill and Decode phases to achieve millisecond-level inference latency [11]. Group 4: Industry Implications - The establishment of the Zhanjiang AI inference cluster represents a shift in the AI infrastructure development logic, moving from merely pursuing computational scale to emphasizing efficiency and cost [13]. - The cluster is expected to provide a significant computational foundation for local industry digital transformation and facilitate the collaborative development of domestic models and chips [14].

Core Viewpoint - The article discusses the rapid rise of OpenClaw, an AI tool that operates 24/7, highlighting its high token consumption and the need for cost reduction in AI applications [1][3]. Group 1: OpenClaw and AI Market Dynamics - OpenClaw has been recognized as a significant software release, surpassing Linux in download speed within three weeks, indicating a major shift in the AI landscape [6]. - The global token consumption has surged by 1000 times due to the widespread use of AI agents for tasks like web searches and data analysis, creating a substantial demand for computational power [8]. - The cost of token usage is a critical barrier to the widespread adoption of AI applications, with a goal to reduce costs significantly over the next five years [3]. Group 2: GPU Innovations and Future Developments - Cloud GPU startup Yuntian Lifa aims to reduce AI application costs by 1 million times by 2030, with plans to release a new GPNPU chip that integrates GPU and NPU capabilities [3]. - The first generation of super node P chips is set to launch in 2026, targeting performance comparable to NVIDIA's Hopper architecture, while the D chips will focus on ultra-low latency inference [4]. - The next-generation architecture, Vera Rubin, will optimize for intelligent agent AI constraints, addressing core issues like long context processing [9].

Company and Industry Summary Industry: Cloud Computing and AI Services Key Points - The ongoing transformation in cloud computing services is compared to the significant changes seen in mobile payments a decade ago, indicating a strong shift towards cloud computing services driven by demand for computational power [1] - Recent developments include D.S. releasing new expectations and granting early access to domestic chips for D.S. V4, marking a notable shift in strategy [1] - China's AI model API call volume has surpassed that of the United States for the first time, with four major models ranking among the top five globally, highlighting the rapid growth and competitiveness of China's AI sector [1] - There is an anticipated shortage of computational power and a wave of price increases expected in the market, suggesting a tightening supply-demand dynamic [1] - The investment focus should be on companies with GPU content, particularly those with domestic offerings showing marginal changes, while those containing NVIDIA products are becoming increasingly scarce [1] - The use of agents in AI is moving towards a more equitable distribution, but access will still require financial investment, indicating a shift in the cost structure of AI services [1] Company Recommendations - The report emphasizes the importance of identifying "computational power partners" and acknowledges operational challenges and power shortages faced by companies [2] - Key recommendations include focusing on Alibaba Group and its affiliates, as well as domestic GPU manufacturers such as Haiguang and Cambricon [2] - Other companies to watch include those in the Huawei ecosystem, ByteDance affiliates, and various tech firms like Wangsu Science & Technology, Dongfang Guoxin, and others, indicating a broad spectrum of investment opportunities in the sector [2]

Core Viewpoint - The dramatic market reaction following Nvidia's record earnings report highlights a significant shift in AI model usage, with China's AI model call volume surpassing that of the U.S. for the first time, leading to a surge in domestic computing power demand and a reevaluation of computing value distribution in the capital markets [1][2][4][5][19]. Group 1: Nvidia's Market Reaction - Nvidia's stock fell by 5.5% on February 26, resulting in a market cap loss of nearly $260 billion (approximately 1.77 trillion RMB) [2][6]. - Despite reporting a record Q4 revenue growth of 73% to $68.1 billion, the market reacted negatively, indicating a shift in focus from short-term performance to long-term sustainability concerns regarding AI capital expenditures [6][25]. - The decline in Nvidia's stock also affected other chip manufacturers, including Broadcom, AMD, and TSMC, which saw varying degrees of stock price drops [6]. Group 2: Surge in Chinese Market - On February 27, A-share and Hong Kong markets experienced a surge in stocks related to computing power leasing, cloud computing, and electricity, with notable gains such as 20% for CloudWalk Technology and 19.91% for Jiawei New Energy [2][10][11]. - The OpenRouter platform reported that during the week of February 9-15, China's AI model call volume reached 4.12 trillion tokens, surpassing the U.S. volume of 2.94 trillion tokens for the first time [19][22]. - By February 16-22, China's call volume further increased to 5.16 trillion tokens, marking a 127% growth within three weeks [19]. Group 3: Shift in Computing Power Demand - The efficient architecture of Chinese models is reducing reliance on high-end GPUs, leading to exponential growth in domestic computing power demand [5][27]. - The "Mixture-of-Experts" (MoE) architecture used by many Chinese models significantly lowers inference costs, allowing for a substantial increase in token usage without a corresponding increase in GPU demand [25][26]. - The cost of processing tokens with Chinese models is significantly lower compared to their U.S. counterparts, with prices as low as $0.3 per million tokens compared to $5 for foreign products [27]. Group 4: Future Outlook - Analysts predict that the demand for domestic computing power will continue to grow exponentially, with a projected compound annual growth rate of 330% for China's token consumption from 2025 to 2030, leading to a 370-fold increase in just five years [27][28]. - The success of Chinese AI models in the global market is expected to validate their performance and cost competitiveness, potentially expanding the domestic computing power market beyond just serving local giants to a global audience [28].

Group 1 - Lito Electronics and Chengdi Xiangjiang reached the daily limit increase [1] - Yuntian Lifey surged over 15% [1] - Shunwang Technology and Capital Online both increased by over 10% [1] - Yunsai Zhili and Wangsu Technology, along with Hongjing Technology, also experienced gains [1]

Market Overview - The total market turnover reached 2.51 trillion yuan, a decrease of 512 billion yuan compared to the previous trading day [1] - The small metals, rare earths, and non-ferrous metals sectors experienced significant growth, while the glass fiber, semiconductor equipment, PCB, and components sectors declined [1] Top Gainers - YN Holdings (豫能控股) saw a rise of 9.98% to 13.34 yuan with a turnover rate of 8.39%, marking its seventh consecutive increase [2] - ST Songfa (*ST松发) increased by 5.00% to 110.09 yuan, with a turnover rate of 2.63%, achieving six increases in eight days [2] - Tongyuan Aluminum (童源铝业) surged by 10.00% to 40.60 yuan, driven by rising tungsten prices and positive performance expectations, marking five increases in seven days [2] - Jiang Aluminum Equipment (江铝装备) rose by 10.02% to 21.53 yuan, with a turnover rate of 9.92%, following a capital increase to acquire tungsten and aluminum assets [2] - Jinzhengda (金正大) increased by 9.93% to 3.10 yuan, attributed to rising fertilizer prices and phosphate mine construction, marking five increases in seven days [2] Market Trends - High-performing stocks included YN Holdings with seven consecutive increases, ST Songfa with six increases in eight days, and several others like Jiang Aluminum Equipment and Jinzhengda with multiple increases [3] - The top net purchases on the daily leaderboard were by Tuo Wei Information, Huasheng Tiancai, and Yuntian Lifu, with net purchases of 5.67 billion yuan, 3.93 billion yuan, and 3.46 billion yuan respectively [4] Sector Performance - The PCB sector faced a significant downturn, impacting companies like Huasheng Tiancai, which saw a drop of 6.74% due to profit-taking after a substantial 98.42% increase over the previous 20 trading days [17] - The environmental protection sector, represented by Qidi Environment, experienced a slight increase of 1.44% amid expectations of restructuring and legislative support for environmental laws [21] Institutional Activity - The top net purchases by institutional investors included Capital Online, Hunan Gold, and Fuling Electric, with net purchases of 1.91 billion yuan, 1.57 billion yuan, and 1.53 billion yuan respectively [7] - Conversely, the top net sales were by Honghe Technology, Qidi Environment, and Hongsheng Huayuan, with net sales of 1.8 billion yuan, 1.33 billion yuan, and 1.07 billion yuan respectively [6] Notable Stocks - Tuo Wei Information's stock price increased by 10.00% to 36.52 yuan, with a turnover rate of 8.17% and a total transaction volume of 33.62 billion yuan [25] - Hunan Gold also saw a 10.01% increase to 37.70 yuan, with a turnover rate of 10.04% and a total transaction volume of 56.73 billion yuan [25]