英伟达产业链观点更新 20251228 摘要 国产替代加速：国内设备厂商能力提升，合作范围扩大，推动半导体设 备如量检测、涂胶显影及空白掩模板等领域的国产替代率显著提高，尤 其在半导体级别应用中潜力巨大。 算力需求激增：下游应用如手机、智能终端推动算力需求，液冷、电源 系统等配套设施成为关键。太空算力、新型电源结构等新兴途径为中国 在 AI 领域提供竞争优势。 海外算力产业链催化：2026 年第一季度海外算力产业链将迎来多重催 化，英伟达 COWS 流片产能预计 2026 年增长 60%-70%，2027 年增 长 50%-60%，谷歌 TPU 出货量预计 2027 年接近千万级别，同比增长 近 100%。 中际旭创盈利预测：市场预期中际旭创 2026 年利润约为 350 亿至 400 亿人民币，若英伟达和谷歌在 2027 年保持高速增长，其利润可能翻倍 至 700 亿至 800 亿人民币，估值有望达到 1 万亿至 1.2 万亿元。 大模型关注焦点：市场关注 OpenAI 的 GPT 和 XAI 的 Grok 大模型，若 基于 B 系列芯片训练的大模型在 2026 年发布后表现超预期，将验证 Scaling ...

Investment Rating - The report indicates a positive outlook for the industry, suggesting a "Buy" rating based on expected growth exceeding the market by over 15% in the next 3-6 months [40]. Core Insights - The competition in large models is intensifying, with significant advancements in capabilities, particularly with Google's Gemini 3 and OpenAI's GPT-5.2, which highlight the potential economic value of large models [1][14]. - The demand for AI applications is accelerating, particularly in inference, as evidenced by the rapid increase in token usage for ByteDance's Doubao AI assistant [2][30]. - The "14th Five-Year Plan" emphasizes the development of strategic emerging industries and future industries, indicating a clear direction for investment in AI and computing infrastructure [3][35]. Summary by Sections Large Model Competition - Major models are continuously iterating, with Gemini 3 showing significant improvements in reasoning and multimodal capabilities, achieving scores of 37.5% and 45.8% in key benchmarks [11][12]. - The transition to the Blackwell architecture is expected to enhance model training capabilities significantly by 2026, indicating that the progress in model capabilities is not yet at a bottleneck [24][26]. Acceleration of AI Application - ByteDance's Doubao AI assistant has transformed mobile interaction, with daily token usage skyrocketing from 16.4 trillion to over 50 trillion in less than a year, reflecting a robust growth in inference demand [2][30]. - NVIDIA's collaboration with Groq, a startup specializing in inference technology, signifies a strategic move towards enhancing inference capabilities, with Groq's LPU architecture designed for high efficiency and low latency [31][34]. Strategic Planning and Industry Layout - The "14th Five-Year Plan" outlines support for strategic emerging industries, including aerospace, quantum technology, and AI, while promoting the construction of new infrastructure for computing power [3][35]. - The report highlights the importance of building a robust ecosystem for emerging industries, focusing on innovation and the application of new technologies [35]. Related Investment Targets - Key investment targets in computing power include companies like Cambricon, Hygon, and Semiconductor Manufacturing International Corporation, while AI agents include major players like Google, Alibaba, and Tencent [4][36]. - The report also identifies potential investments in autonomous driving and military AI sectors, with companies such as Xpeng Motors and Tsinghua Tongfang listed as notable players [5][38].

Group 1: Industry Insights - The competition in large models remains intense, with the Scaling Law still effective. Google's Gemini 3 has made significant advancements in foundational reasoning and multimodal capabilities, while OpenAI's GPT-5.2 emphasizes the potential of large models in creating economic value [1][11][14] - Meta is actively developing two heavyweight AI models, Mango for image and video processing, and Avocado to enhance programming capabilities, indicating a strong commitment to AI development [1][15] - The Chinese open-source model DeepSeek-V3.2 is approaching the performance of top closed-source models, showcasing innovations in sparse attention (DSA), high post-training ratios, and large-scale synthetic data [1][16][18] Group 2: AI Application Acceleration - ByteDance released the Doubao AI mobile assistant, which allows for cross-application autonomous operations, marking a significant evolution in mobile interaction methods [2][26] - The daily token usage of the Doubao model surged from over 16.4 trillion in May 2025 to over 50 trillion by December 2025, reflecting a rapid increase in inference demand [2][28] - NVIDIA's collaboration with Groq, a startup specializing in inference chips, highlights a strategic move towards enhancing inference capabilities while maintaining its dominance in training power [2][29][30] Group 3: Policy and Future Industry Layout - The "14th Five-Year Plan" emphasizes support for strategic emerging industries such as aerospace, quantum technology, and AI, indicating a clear direction for future industrial development [3][42] - The plan also calls for proactive infrastructure development, including information communication networks and integrated computing networks, reinforcing the importance of computational power in the AI era [3][42] Group 4: Related Companies - Key players in computing power include Cambrian, Haiguang Information, and Zhongke Shuguang, among others, indicating a diverse landscape of companies involved in AI and computing infrastructure [4][43] - Companies involved in AI agents include Google, Alibaba, Tencent, and others, showcasing a broad spectrum of firms engaged in AI development [5][44] - In the autonomous driving sector, companies like Jianghuai Automobile and Xiaopeng Motors are notable participants, reflecting the industry's growth [6][45]

2025年12月，在腾讯科技HiTechDay上，以《模型再进化：2025，智能重新定义世界》为主题的圆桌论坛，正是围绕大模型进化的深度、维度、效率三条 线索展开。 华中师范大学人工智能教育学部助理教授熊宇轩为嘉宾主持，三位嘉宾北京智源人工智能研究院院长王仲远、面壁智能联合创始人、首席科学家刘知远、 峰瑞资本投资合伙人陈石分别从各自的领域，解读2025对于大模型进化的深入观察。 王仲远指出，大模型的进化正在经历"从Learning from Text到Learning from Video"的质变。视频数据中蕴含了丰富的时空信息与动态交互线索，为模型学 习物理世界动态演变规律提供了关键的数据来源，同时也是当前最容易规模化获取的一类多模态数据，是AI"从数字世界迈向物理世界"的关键桥梁，也为 具身智能（Embodied AI）的爆发提供了构建"世界模型"的底座。 刘知远提出的"密度法则"（Densing Law）认为，如同芯片摩尔定律，AI的未来在于不断提升单位参数内的"智能密度"。他大胆预言，未来的算力格局将 是"云端负责规划，端侧负责做事（执行）"，到2030年，我们甚至有望在端侧设备上承载GPT-5级别的 ...

2025 年 12 月初，圣地亚哥 NeurIPS 大会。 Geoffrey Hinton（神经网络奠基人、2024年诺贝尔物理学奖得主）与Jeff Dean（Google首席科学家、 Gemini模型联合负责人、TPU架构师）的炉边对谈，成为这场大会的重要时刻。 对话聚焦一个关键问题： 现代 AI 为什么能从实验室走向数十亿用户？ 从 AlexNet 在学生卧室的两块 GPU 上训练，到 Google 在餐巾纸上算出TPU需求；从学术圈的小众实 验，到支撑全球亿级应用的基础设施。 这是一次对 AI 工业化进程的系统性复盘。 他们给出的答案是：现代 AI 的突破从来不是单点奇迹，而是算法、硬件、工程同时成熟后的系统性涌 现。强算法必须与强基础设施结合，才能真正走向规模化。 看清这条路径，你就能理解AI为什么是今天这个样子。 第一节｜AI的突破，起于一块GPU板 Geoffrey Hinton 说，现代 AI 真正的转折，不在某篇论文里，而是在他学生 Alex 的卧室里：两块 NVIDIA GPU 板，插在父母家电脑上，训练图像识别模型。电费，还是家里人掏的。 那是 2012年 ，ImageNet 比赛。 别人 ...

AI端侧，实则是一个庞大而复杂的产业链生态，是大模型装入消费电子的外设+应用场景+数据挖掘+数 字资产+信用扩张周期。当下，Qwen、智谱等中国模型已经逐渐被全球开发者所认可。而所谓的外设， 就是靠着强大的制造能力而风靡全球的"中国制造"，比如消费电子、电动车、无人机、机器人等等。AI 端侧的核心意义，就在于解决AI基建那巨量资本开支的变现问题，并构建完整的AI端侧产业链生态闭 环。 现在西大在力推谷歌的"全栈自研"模式：试图将芯片-软件-模型-应用的链条全部跑通，形成自身的竞争 优势。而东大，则更应当构建独属于自己的AI战略。中国人的优势就是制造Power，机器人、无人机、 智能外设、电动车，这些工业制造领域才是中国人大展身手的主战场，软件只是工具。开源的软件让全 世界的用户加入AI生态，然后把这些生态装入物美价廉的硬件（AI Agents）中，最终再把这些硬件卖 给全世界…… 而AI生态催生出的应用场景和数据资源，将形成巨大的飞轮效应，让AI所创造的巨额财富，最终沉淀 在中国的土地上…… 炒股就看金麒麟分析师研报，权威，专业，及时，全面，助您挖掘潜力主题机会！ 来源：刘煜辉的高维宏观 谈及明年展望，刘博对 ...

Core Insights - The AGIX index aims to capture the beta and alphas of the AGI era, which is expected to be a significant technological paradigm shift over the next 20 years, similar to the impact of the internet [2] - The "AGIX PM Notes" serves as a record of thoughts on the AGI process, inspired by legendary investors like Warren Buffett and Ray Dalio, to witness and participate in this unprecedented technological revolution [2] Market Performance - AGIX recorded a weekly performance of 6.00%, a year-to-date return of 26.73%, and a return of 74.56% since 2024 [4] - In comparison, QQQ, S&P 500, and Dow Jones had year-to-date returns of 21.13%, 16.45%, and 12.16% respectively [4] Sector Performance - The application sector saw a weekly performance of 2.20% with an index weight of 33.62% - The semi & hardware sector had a weekly performance of 1.76% with an index weight of 24.22% - The infrastructure sector recorded a weekly performance of 2.08% with an index weight of 37.19% [5] AI Industry Developments - Ilya's recent interview sparked significant market discussion, highlighting concerns about model training stagnation while also noting advancements in Google's Gemini 3 capabilities [9][10] - The AI industry is transitioning from a research phase to a focus on productization and optimization, with Google leveraging its TPU technology for enhanced performance [10] - The future of AI may not be dominated by a single model but rather by productization capabilities and external factors such as distribution and ecosystem [11] Investment Trends - The AI startup financing landscape remains robust, with 49 companies securing over $100 million in single rounds by November, matching the total for 2024 [17] - Major investments include Anysphere's $2.3 billion funding round and OpenAI's record $40 billion financing, indicating a growing concentration of capital in the AI sector [17] Corporate Actions - ServiceNow is in talks to acquire cybersecurity startup Veza for over $1 billion, which would enhance its identity management capabilities [19] - Zscaler reported strong Q1 results but saw its stock drop over 7% due to a conservative outlook, reflecting investor expectations for tech company growth [19]

Core Insights - The article discusses the emerging field of embodied intelligence, highlighting the importance of data generation rates and physical accuracy in developing effective world models for AI systems [2][3][32]. Group 1: Embodied Intelligence Developments - Tesla's Shanghai Gigafactory has announced the mass production of Optimus 2.0 and opened a developer platform to address data isolation issues through ecosystem collaboration [2]. - NVIDIA introduced a comprehensive physical AI solution at the SIGGRAPH conference, aiming to tackle the shortage of real-world data by generating high-quality synthetic data [2]. Group 2: Efficiency Law and Scaling Law - The article introduces the concept of Efficiency Law, which posits that the performance of embodied intelligence models is significantly influenced by the rate of high-quality data generation (r_D) [7][21]. - Scaling Law, previously observed in large language models, faces challenges in the embodied intelligence domain due to the lack of a data paradigm that supports it [6][7]. Group 3: World Models and Physical Accuracy - Current video-based world models focus on visual realism but often lack an understanding of physical laws, leading to inaccuracies in simulating real-world dynamics [9][10]. - The necessity for world models to adhere to physical accuracy is emphasized, as they must enable agents to follow physical laws for effective learning and decision-making [10][11]. Group 4: Generative Simulation World Models - The GS-World model integrates generative models with physical simulation engines, allowing for the generation of environments that adhere to physical laws, thus overcoming the limitations of traditional video-based models [13][14]. - GS-World serves as a foundation for a new learning paradigm, enabling agents to learn through interaction in a physically accurate environment [18][19]. Group 5: Engine-Driven Learning Paradigm - The transition from data-driven to engine-driven learning is highlighted as a fundamental shift, allowing agents to autonomously generate and interact within a simulated world [24][25]. - This new paradigm enhances learning efficiency, generalization capabilities, and interpretability by enabling agents to learn from their own generated experiences rather than relying solely on external data [24][25]. Group 6: Applications and Future Directions - GS-World has significant potential applications, including in reinforcement learning, where it can facilitate high-fidelity strategy validation and optimization [15][16]. - The article concludes with a call for industry and academic collaboration to advance the development and deployment of embodied intelligence technologies based on the GS-World model [33].

Core Insights - The article discusses the evolution of AI from "cognition" to "action," highlighting the transition of Tang Jian from academia to industry, particularly in the fields of autonomous driving and embodied intelligence [1][2] - Tang Jian emphasizes the importance of experience-driven control methods over traditional mathematical modeling in complex environments, suggesting that AI systems can learn from historical data to make effective decisions [4][5] - The concept of a "world model" is introduced as essential for embodied intelligence, enabling robots to understand and predict their environment, thus enhancing their operational capabilities [13][14] Summary by Sections Transition from Academia to Industry - Tang Jian, a former tenured professor, shifted focus to practical applications of AI in industry, particularly in autonomous driving and robotics [1][3] - His experience in various companies, including Didi and Midea, has informed his approach to AI-driven system control [3][6] Experience-Driven Control - The article outlines the difference between traditional control methods and experience-driven approaches, with the latter relying on data and historical experiences rather than precise mathematical models [4][5] - This experience-driven philosophy is evident in autonomous driving applications, where end-to-end control merges perception, planning, and control into a single learning process [6][7] Embodied Intelligence and World Models - Tang Jian argues that embodied intelligence presents a higher complexity than autonomous driving, requiring robots to manage multiple joints and navigate dynamic environments [7][8] - The world model is described as a critical component for robots to understand and interact with the physical world, enabling them to perform tasks that require nuanced understanding and adaptability [14][15] - The article highlights the need for a world model to facilitate the development of robots that can generalize across various tasks and environments, which is crucial for their deployment in real-world scenarios [21][22] Future Directions and Challenges - The discussion includes the potential for world models to achieve a "DeepSeek moment" in embodied intelligence, drawing parallels to breakthroughs in AI performance under limited resources [9][10] - Tang Jian acknowledges the current limitations in data and model architecture, indicating that further iterations and improvements are necessary for the field to progress [2][13] - The article concludes with the assertion that the world model is not just a technical choice but a fundamental requirement for the advancement of embodied intelligence [13][22]

Core Insights - GPT-5 has been released, but it does not represent a significant step towards Artificial General Intelligence (AGI) [1] - The launch event revealed several issues, including presentation errors and reliance on debunked theories, which highlighted weaknesses in the Transformer architecture [1] - Despite these shortcomings, GPT-5 is still considered a competent AI product, and OpenAI plans to implement aggressive commercialization strategies in key sectors [1] Technical Development - The development of GPT-5 faced various technical bottlenecks, leading to the choice of a specific architecture to overcome these challenges [1] - The limitations of the Scaling law have been encountered, raising questions about future technological pathways for AI advancement [1] Commercial Strategy - OpenAI aims to rapidly establish a presence in three main application areas: education, healthcare, and programming [1] - The company's approach suggests a focus on leveraging GPT-5's capabilities to solidify its market position [1]