宗建树 陈耀文 SAC：S0490520030004 SAC：S0490525070002 SFC：BUX668 $${\mathrm{iii}}\{8\}{\mathrm{iii}}\{8\}{\mathrm{iii}}\{8\}{\mathrm{iii}}\{8\}{\mathrm{iii}}\{8\}{\mathrm{iii}}\{8\}$$ 丨证券研究报告丨 行业研究丨点评报告丨软件与服务 [Table_Title] AI 产业跟踪：GPT-5.2 发布，关注大模型迭代进 展 报告要点 [Table_Summary] 当地时间 12 月 11 日，OpenAI 正式发布新一代大模型 GPT-5.2，主打通用智能、编码能力和 长上下文任务处理，作为近期连续版本更新中的又一项核心升级。此次，OpenAI 发布"迄今为 止功能最强大的专业知识工作模型系列"——GPT-5.2，在提升性能的同时，将重心逐步转移 至其模型货币化，或将加速大模型的商业化落地。建议关注：1）国内大模型厂商；2）国内大 型云厂商；3）垂类场景 Agent 厂商；4）国产算力产业链。 分析师及联系人 [Table_Author] %% ...

Report Summary 1. Report Industry Investment Rating No industry investment rating is provided in the report. 2. Core Viewpoints of the Report - In the US, multiple quantitative enhancement products were issued last week, with an increasing issuance speed at the end of the year. Various asset classes in US ETFs maintained inflows, and alternative strategies such as long - short equity performed well. US domestic stock - type mutual funds still faced significant redemption pressure, while bond funds had a slight inflow [2]. 3. Summary by Directory 3.1 US ETF Innovation Products: Multiple Quantitative Enhancement Products Issued - Last week, 43 new products were issued in the US, including 6 individual stock leveraged products and 3 digital currency - related products. One product combined crude oil and Bitcoin with 2x leverage, and Simplify's US stocks + futures strategy also had a 1:1 investment ratio. Motley Fool issued 3 single - factor ETFs, each holding about 150 stocks [5][6]. - BlackRock's quantitative team issued an alternative product, and NEOS issued a long - short equity product. Hedgeye's 130/30 product also adopted a long - short strategy. Global X issued a gold miners ETF, Franklin Templeton issued a small - cap enhanced ETF, and Sterling Capital's stock option product used a quantitative stock - selection strategy [7]. - Columbia issued 6 ETFs, 3 bonds and 3 stocks. The stock products mainly used a quantitative enhancement strategy with semi - annual rebalancing [8]. 3.2 US ETF Dynamics 3.2.1 US ETF Fund Flows: All Asset Classes Maintained Inflows - In the past week, US ETF inflows remained above $40 billion, and domestic stock products had inflows of over $30 billion. There was a significant difference in fund flows between BlackRock's S&P 500 ETF (outflow) and Vanguard's products (inflow). Russell 2000 and high - yield bond ETFs had inflows, indicating a relatively high risk appetite [2][9]. - S&P 500 ETFs had significant recent fund fluctuations, Russell 2000 ETFs had continuous inflows, and gold also returned to an inflow state [13]. 3.2.2 US ETF Performance: Alternative Strategies such as Long - Short Equity Performed Well - Many long - short equity products were issued last week. In the past two years, products replicating futures and combining multiple hedge fund strategies have been increasing. Among the top ten alternative strategy products in the US, State Street's multi - strategy product and Convergence's long - short equity product performed best [14]. 3.3 Recent Fund Flows of US Ordinary Public Offering Funds - In October 2025, the total amount of non - money public offering funds in the US was $23.7 trillion, an increase of $0.22 trillion from September. The S&P 500 rose 2.27% in October, and the scale of domestic stock - type products increased by 0.9%, but the redemption pressure was still high. - From November 25th to December 3rd, domestic stock funds in the US had outflows of over $15 billion. Hybrid products had continuous outflows, while bond funds had a slight inflow [15].

Summary of Conference Call on Technology Sector Industry Overview - The conference call focused on the AI industry, particularly the development trends of AI models and their applications in various sectors, including storage, computing power, and PCB markets [1][2][3]. Key Points and Arguments AI Model Development - The trend in AI model development emphasizes enhancing intelligence density, stability, and cost efficiency under unit computing power, with domestic companies like Deepseek and Tongyi Qianwen excelling in cost optimization and open-source initiatives [1][2]. - Recent releases include Google's Gemini 3, which restructured search functionalities, and OpenAI's GPT-5.2, showing significant improvements over previous versions [2]. Productization of AI Models - AI models are increasingly being productized, with examples such as Google's Gemini 3 and the Doubao app, which utilizes operating system-level permissions for cross-app price comparisons [1][3][4]. - The Robot Taxi industry is projected to reach a market size of 100 billion RMB by 2035, with a penetration rate of 10%, indicating a shift from conceptual discussions to practical implementations [3][14]. Storage and Computing Power Trends - The AI industry is driving growth in upstream storage and computing power, with DRAM capital expenditures expected to exceed NAND, and DDR4 prices strengthening due to supply structure adjustments [1][6][7]. - TrendForce predicts DRAM capital expenditures could reach $61 billion by 2026, a 14-15% increase, while NAND is expected to grow by 5% to over $22 billion [7]. PCB Market Insights - The PCB market is currently chaotic, but investment opportunities in Google's ASIC chain are seen as stronger than those in NV chains, with recommendations for companies like ShenNan, Huadian, and Dongshan [11]. - The introduction of mid-plane designs in PCBs is expected to enhance system stability and drive demand and prices upward [9][10]. End-User Market and Future Products - The end-user market is expected to see innovative products such as OpenAI hardware and Apple AI glasses, with specific companies recommended for investment [12]. - The upcoming Siri iteration in 2026 is anticipated to enhance Apple's valuation [12]. Robot Taxi Industry Analysis - The Robot Taxi model focuses on providing a consistent user experience rather than merely competing on price, with local governments showing positive attitudes towards its development [13][14]. - The industry has transitioned from conceptual hype to practical validation, with companies like Waymo and domestic players like Xiaoma Zhixing and Wenyuan Zhixing making significant progress [15]. Autonomous Driving Technology - L4 autonomous driving technology is expected to have a short-term advantage, while L2+ technology companies like Xiaopeng Motors and Horizon Robotics may also see valuation opportunities [16][17]. Additional Important Insights - The emphasis on product design and user experience in AI applications is crucial for future competitiveness [5]. - The need for substantial storage resources in both training and inference phases of AI models highlights the importance of selecting the right stocks in the semiconductor sector [6].

Report Industry Investment Rating No information about the report industry investment rating is provided in the content. Core Viewpoints of the Report - The issuance speed of US ETFs at the end of the year has increased again, with multiple quantitative enhancement products being issued [2][7]. - The capital inflow of US ETFs has remained above $40 billion, and the risk appetite of capital has remained at a high level [2][13]. - Stock long - short and other alternative strategies of US ETFs have performed well [2][19]. - The redemption pressure of US non - money mutual funds in October 2025 was still high, and domestic stock funds and hybrid products have continued to experience outflows recently, while bond funds have seen a slight inflow [2][20]. Summary by Relevant Catalogs 1. US ETF Innovation Products: Multiple Quantitative Enhancement Products Issued - Last week, 43 new products were issued in the US, including 6 individual stock leverage products and 3 digital currency - related products [2][7]. - Motley Fool issued 3 single - factor ETFs, each holding about 150 stocks [9]. - BlackRock's quantitative team, NEOS, Hedgeye, Global X, Franklin Templeton, Sterling Capital, and Columbia all issued different types of ETFs last week, with many using quantitative strategies [10][11]. 2. US ETF Dynamics 2.1 US ETF Capital: All Types of Assets Maintain Inflows - In the past week, the inflow of US ETFs has remained above $40 billion, and the inflow of domestic stock products has exceeded $30 billion [2][13]. - The S&P 500 ETF of BlackRock continued to have the largest outflow, while the products of Vanguard had a large - scale inflow of over $40 billion, with a capital flow difference of over $80 billion between the two. The Russell 2000 and high - yield bond ETFs had inflows [2][15]. 2.2 US ETF Performance: Stock Long - Short and Other Alternative Strategies Perform Well - Many stock long - short products were issued last week, and products combining futures replication and multiple hedge fund strategies have been increasing in the past two years. Among the top ten alternative strategy products in the US, the multi - strategy product of State Street and the stock long - short product of Convergence performed the best [2][19]. 3. Recent Capital Flows of US Ordinary Mutual Funds - In October 2025, the total amount of US non - money mutual funds was $23.7 trillion, an increase of $0.22 trillion compared to September. The scale of domestic stock products increased by 0.9%, but the redemption pressure was still high [2][20]. - From November 25th to December 3rd, the outflow of US domestic stock funds remained above $15 billion. Hybrid products have continued to experience outflows recently, while bond funds have seen a slight inflow [2][20].

深圳市沃尔核材（002130）股份有限公司(简称：沃尔核材，002130.SZ)于6月16日所递交的港股招股书满6个月，于12月16日失效，递表时中信建投 （601066）国际、招商证券国际为其联席保荐人。 招股书显示，沃尔核材是高速数据通信与可替代能源电力传输综合解决方案提供商，亦是制造及销售高速数据通信重要元器件高速铜缆市场的领军企业。有 关产品实现算力中心中功能模块之间的高速连接，加快群集内数据传输效率的同时确保最佳能源消耗和坚实的可靠性。有效促进了高质量基础设施的快速部 署，推动大模型在不同行业中的广泛应用。 ...

作者丨 许华哲 编辑丨具身智能之心 本文已经得到许华哲博士的授权，未经允许，不得二次转载。 点击下方 卡片 ，关注" 具身智能之心 "公众号 >> 点击进入→ 具身 智能之心 技术交流群 昨天看到了许华哲老师在社交媒体上的分享，关于数据、量产、本体和场景。类似的观点，今年IROS圆桌期间，许博也站在智能第一性原理上，将具身的未来发展 方向划分为欲望、先验和经验三个模块。 欲望。 在做智能体的时候，无论是物理的还是虚拟的，总觉得现在机器学习没有自己的学习欲望。我们可以设想一下，能不能给机器人一种自己的欲望？ 经验。 经验是完成世界最终闭环的一种手段。有一天，在家里面看到一位维修师傅就是帮我们修煤气灶，他踩在一个梯子上拧一个东西，整个身体造型极为扭曲， 但他仍可以完美控制重心保持平衡，并且手上还可以做非常精细的操作。 ★ 这种思想也贯穿在后续的研发和学术探索上。 回想起几年前，我们还在讨论机器人什么时候能全地形走路，后来发现这个话题变成了"跑酷"、"跳舞"、"篮球"。这个变化速率让我知道这个事儿已经成了，如果 明年可以攀岩我并不吃惊。 但这极快的变化速率又显得格外不协调，因为我没在任何地方看到人形机器人真正服务人 ...

Core Viewpoint - The automotive industry is facing a significant shift towards software-related issues, with software defects now accounting for a substantial portion of vehicle recalls, indicating a systemic risk that must be addressed through enhanced cybersecurity measures and a holistic approach to safety [7][15][19]. Group 1: Software-Related Recalls - In 2024, the total number of vehicles recalled globally due to software issues is projected to reach 13.4 million, which is over four times the number from 2023, representing 46% of all recalls [7][15]. - The ratio of recalls due to software defects is now nearly equal to that of traditional mechanical design defects, highlighting the growing importance of software safety in the automotive sector [7][15]. Group 2: Cybersecurity and Systemic Risks - The transition towards "new four modernizations" in the automotive industry, including electrification and connectivity, has expanded the attack surface for vehicles, necessitating urgent exploration of cybersecurity measures [7][9]. - Experts emphasize the need for a comprehensive, system-wide approach to automotive cybersecurity, integrating security from the ground up in the development process rather than as an afterthought [10][12]. Group 3: AI and Future Challenges - AI is seen as both an enabler and a potential source of unforeseen challenges in automotive cybersecurity, with the rapid evolution of AI technologies posing risks that are not yet fully understood [18][19]. - The integration of AI into automotive systems requires a reevaluation of existing security frameworks, as traditional methods may not adequately address the complexities introduced by AI [54][56]. Group 4: OTA Security Measures - Companies are implementing various strategies to ensure the security of Over-The-Air (OTA) updates, including dual backup systems and real-time user feedback during the update process [40][44]. - The balance between user experience and safety is critical, with companies prioritizing safety over convenience when necessary [41][44]. Group 5: Collaboration and Testing - Collaboration with third-party security firms for testing and validation is common, as companies recognize the need for external expertise in identifying vulnerabilities [50][51]. - Continuous testing and updates are essential for maintaining security throughout the vehicle's lifecycle, akin to regular health check-ups for humans [55].

Core Viewpoint - The article discusses the advancements in the development of MindGPT, a multimodal cognitive model designed to enhance human-machine interaction in smart vehicles, emphasizing its capabilities in perception, understanding, and interaction [2][20][39]. Group 1: Technology and Model Architecture - MindGPT is built on a self-developed TaskFormer structure, which has been recognized for its performance in industry evaluations [2][35]. - The model incorporates multimodal perception capabilities, allowing it to process audio and visual data simultaneously, enhancing user interaction through features like voice recognition and gesture control [29][30]. - The architecture supports a complete agent capability, integrating perception, planning, memory, tools, and action [35][36]. Group 2: Training and Performance - The training strategy focuses on 15 key areas relevant to in-car scenarios, utilizing self-supervised learning and reinforcement learning from human feedback (RLHF) to cover over 110 domains and 1,000 specialized capabilities [3][35]. - The training platform, Li-PTM, achieves training speeds that are significantly faster than industry standards, with SFT phase speeds over three times better than the best open-source capabilities [46][47]. - The model's inference engine, LisaRT-LLM, has been optimized for performance, achieving a throughput increase of over 1.3 times compared to previous models under high concurrency [5][53]. Group 3: User Interaction and Experience - MindGPT aims to create a natural interaction experience by allowing users to communicate with the vehicle using simple commands and gestures, reducing the complexity of user input [10][32]. - The system is designed to understand and remember user preferences, providing personalized interactions based on historical conversations [36][39]. - The integration of advanced AI technologies aims to enhance emotional connections between users and their vehicles, creating a more immersive experience [14][18].

Group 1 - The event titled "The Breakthrough Path of Domestic GPUs in the Era of Large Models" will take place on December 18, from 14:00 to 16:00 at Tsinghua University's School of Economics and Management [1] - The event is organized by Tsinghua University's Blockchain Finance Research Center, Tsinghua x-lab, and QLChain Academy, focusing on the intersection of technology and business innovation [1][3] - The speaker, Yan Yan, is the Commercial Product Solutions Director at Beijing Xiwang Chip Intelligent Technology Co., with 20 years of industry experience, particularly in AI technology applications across various sectors [2] Group 2 - Tsinghua University's Blockchain Finance Research Center was established in April 2018, aiming to leverage technological advantages to promote business innovation and enhance understanding of blockchain in traditional industries [2] - Tsinghua x-lab, founded in 2013, serves as an educational platform for discovering and cultivating innovative talents, supporting China's innovation-driven development strategy [3] - QLChain Academy is a platform initiated by Tsinghua alumni for blockchain technology exchange, investment, and resource interaction, linking resources within and outside the university [3]

Core Insights - The "Tongyi" team has announced significant upgrades to its Fun-CosyVoice3 model, including a 50% reduction in initial latency and a doubling of accuracy for mixed Chinese-English speech recognition [1][2] - The Fun-CosyVoice3 model is now open-source, featuring zero-shot voice cloning capabilities and supporting local deployment and customization [1] - The Fun-ASR-Nano model has been introduced, with a reduced parameter count of 0.8 billion, aimed at lowering inference costs while also being open-source [1] Group 1 - Fun-CosyVoice3 model has achieved a 50% reduction in initial latency, enabling real-time applications such as voice assistants and live dubbing [2] - The word error rate (WER) for mixed Chinese-English speech has decreased by 56.4%, improving accuracy in complex sentences and professional terminology [2] - The model supports 9 common languages and 18 Chinese dialects, with cross-lingual voice cloning capabilities, allowing for high consistency in voice reproduction across different languages [2] Group 2 - The Fun-ASR model has undergone comprehensive upgrades, enhancing robustness in noisy environments and supporting multilingual speech recognition [2] - The first-word latency for the streaming recognition model has been reduced to 160 milliseconds, improving responsiveness in applications [2] - Fun-ASR has been successfully implemented in various scenarios, including DingTalk's "AI Listening" and video conferencing [2]