Core Viewpoint - The global shared mobility market is undergoing a critical transition from human-driven to automated services, exhibiting significant regional differentiation [4][9]. North America Market - The North American ride-hailing market is dominated by Uber and Lyft, creating a stable pricing power. In the Robotaxi sector, Waymo holds a monopoly while Tesla aggressively disrupts the market. Chinese Robotaxi companies face barriers due to a 2025 U.S. Department of Commerce ban on hardware and software, complicating their commercialization path [4][9][16]. European Market - The European regulatory environment is fragmented and stringent, with local automakers lagging in L4 algorithm development. This creates a unique "hybrid model" opportunity, where "U.S./local platforms + Chinese technology" could break through. Uber and Lyft's collaboration with Baidu Apollo indicates that de-branding technology output is a favorable solution for entering the European market [4][9][16]. Middle East Market - The Middle East presents a unique "three highs and one low" characteristic: high customer spending, high policy support, high infrastructure investment, and low energy costs. Gulf countries are eager to reduce oil dependency, viewing autonomous driving as a national strategy. Chinese companies like WeRide and Pony.ai benefit from dual advantages of road rights and licenses, making it an ideal training ground and commercialization area for overseas expansion [4][9][16]. Southeast Asia Market - The Southeast Asian ride-hailing market is large but has low customer spending. Low labor costs may lead to economic challenges for Robotaxi operations. In the short term, large-scale deployment of Robotaxis is not cost-effective, and two-wheeled vehicles remain mainstream. Singapore, with its high labor costs, may achieve Robotaxi commercialization [4][5][9]. Investment Focus - Focus on the L4 RoboX industry chain, prioritizing B-end software over C-end hardware. Recommended stocks include: - Hong Kong stocks: Xpeng Motors, Horizon Robotics, Pony.ai, WeRide, Cao Cao Mobility, and Black Sesame Technology - A-shares: Qianli Technology, Desay SV, and Jingwei Hirain - Downstream application-related stocks from the Robotaxi perspective include integrated models (Tesla, Xpeng Motors), technology providers with revenue-sharing models (Horizon, Baidu, Pony.ai, WeRide, Qianli Technology), and the transformation of ride-hailing/taxi services (Didi, Cao Cao Mobility, Ruqi Mobility, Dazhong Transportation, Jinjiang Online) [6][9]. Regulatory and Market Barriers - The regulatory landscape for Robotaxis abroad features a dual approach of support and regulation. Companies must assume clear accident liability and purchase sufficient liability insurance. Vehicles must have complete data recording capabilities and undergo third-party safety assessments. Operationally, there are restrictions on operational areas, fleet size, and speed [12][14]. Market Size and Growth - The North American shared mobility market is projected to grow significantly, with the total Gross Transaction Value (GTV) expected to reach billions by 2030. The European market also shows substantial potential, albeit with slower conversion rates. The Middle East is characterized by strong government support, while Southeast Asia presents a high-growth potential due to infrastructure gaps [21][22][27]. Pricing Dynamics - Pricing dynamics vary significantly across regions, influenced by local labor costs and regulatory environments. North America has high labor costs, allowing Robotaxis to survive without extreme price reductions. In contrast, Europe faces stringent labor protections that increase operational costs. The Middle East's pricing is shaped by government-led transportation strategies, while Southeast Asia's ultra-low fares are supported by low labor costs [33][34]. Profitability Disparities - Profitability varies significantly across countries, with developed regions showing higher absolute margins per Robotaxi. Revenue per vehicle in China, UAE, UK, and the US is estimated at approximately $40,000, $90,000, $250,000, and $250,000 respectively, with gross margins reflecting these disparities [34][35].

Investment Rating - The report does not explicitly state an investment rating for the industry Core Insights - The competition in AI entry points is intensifying, with major companies increasing their investments. China's AI presence globally has significantly improved, with domestic large models continuously iterating. Despite GPT-5.2 and Gemini 3 Pro leading, Chinese models have effectively altered the North American dominance in the competitive landscape. In the global Top 10, three positions are held by Chinese models, and in the Top 15, there are six Chinese companies. By 2025, China's open-source AI model usage is expected to account for over 70% of the global market [2][11][19] - The demand for inference has surged, with the emergence of o1 class inference models unlocking approximately 10 times the potential of traditional models in terms of inference-time compute. The demand for computing power has shifted from being solely "training-driven" to a dual focus on "training + inference" [2][5][37] - The battle for entry points has evolved beyond mobile devices to OS-level intelligent agents and super apps. By December 24, 2025, ByteDance's AI application Doubao announced daily active users (DAU) exceeding 100 million, while Qianwen App reached over 30 million monthly active users within 23 days of public testing, becoming the fastest-growing AI application globally. Doubao bypasses traditional interfaces, creating an "AI operating system" that directly interacts with super apps like WeChat and Alipay, challenging the rules of the traditional app era [2][44][45] Summary by Sections AI Entry Point Competition - China's AI global presence has significantly improved, with domestic large models continuously iterating. In the global Top 10, three positions are held by Chinese models, and in the Top 15, there are six Chinese companies. By 2025, China's open-source AI model usage is expected to account for over 70% of the global market [2][11][19] - The competition for entry points has evolved beyond mobile devices to OS-level intelligent agents and super apps, with significant user engagement reported for new AI applications [2][44][45] Domestic Chip Breakthroughs - The smart computing center in China is expanding, with a projected compound annual growth rate (CAGR) of 57% from 2020 to 2028, reaching 2,781.9 EFLOPS by 2028. Domestic chip technology is steadily improving, with local cloud service providers accelerating the construction of heterogeneous environments [5][50] - Domestic general-purpose GPUs are upgrading from "usable" to "good," with performance metrics approaching those of leading international models. The production capacity of domestic chip manufacturers like SMIC is continuously increasing, providing solid support for domestic AI chip production [5][53][54] Supply and Demand Dynamics - The demand side is characterized by a surge in inference demand as AI applications become more prevalent, while the supply side sees continuous improvements in domestic GPU performance and accelerated adaptation by cloud service providers [5][59] - The AI server market is expected to see a shift towards inference servers becoming the mainstream, with a projected market size of approximately $39.3 billion in 2024, reflecting a year-on-year growth of 49.7% [5][64]

Core Viewpoint - The article discusses the challenges faced by input method applications in the era of AI, highlighting user frustrations with accuracy and excessive advertisements, despite significant investments from major tech companies in enhancing these tools [2][4][14]. Group 1: User Experience Issues - Users are increasingly dissatisfied with input methods, reporting issues such as inaccurate word predictions and excessive advertisements, which detract from the overall user experience [2][4]. - A specific case is mentioned where a long-time user of Sogou Input Method uninstalled the app due to frequent incorrect suggestions, indicating a decline in basic functionality despite advanced AI features [4]. - Complaints about the voice recognition capabilities of input methods have also surfaced, with users noting that corrections often take longer than typing the text directly [4]. Group 2: AI Integration and Competition - Major input method providers, including Sogou, Baidu, and iFlytek, are engaged in a competitive race to integrate advanced AI capabilities into their products, aiming to enhance user experience and functionality [2][8][9]. - The input method market is characterized by a concentrated structure, with leading companies holding a combined market share of 84.4% as of July 2025, indicating a competitive landscape [8]. - Input methods are evolving from simple typing tools to becoming the primary interface for AI interactions, with companies aiming to position their products as essential gateways to AI capabilities [9][10]. Group 3: Commercialization and Privacy Concerns - Input methods face challenges in monetization, struggling with a "high traffic, low value" dilemma, which complicates their ability to generate revenue despite having a large user base [15][16]. - Privacy concerns are paramount, as input methods have been criticized for collecting unnecessary personal information, leading to regulatory scrutiny and the need for companies to adapt their data collection practices [16]. - Companies are implementing features that allow users to choose between different modes of data collection, balancing functionality with privacy protection [16]. Group 4: Future Directions - The future of input methods is seen as a shift towards becoming intelligent agents that can understand user intent and context, moving beyond basic text input to more complex interactions [12]. - Companies are exploring multi-modal input methods that incorporate voice, text, and images, which require sophisticated algorithms and technology to manage effectively [17].

Core Insights - The article discusses the challenges faced by input method applications, particularly focusing on Sogou Input Method, as it approaches its 20th anniversary in 2026. Despite advancements in AI capabilities, user experience has deteriorated, leading to complaints about inefficiency and inaccuracies in word prediction and voice recognition [2][3][5]. User Experience Issues - Users express frustration over the declining accuracy of input methods, with complaints about incorrect word suggestions and excessive advertisements disrupting their experience [3][5][6]. - A long-time user of Sogou Input Method reported issues with common character suggestions, indicating a failure to learn user habits despite repeated corrections [4][5]. AI Integration and Competition - Major input method companies, including Tencent's Sogou, Baidu, and iFlytek, are engaged in a competitive race to integrate advanced AI features into their products, aiming to enhance user experience and functionality [3][9][10]. - The input method market is characterized by a concentrated structure, with leading companies holding a combined market share of 84.4% as of July 2025, indicating a fierce competition for user engagement and AI capabilities [9]. Future Development Goals - Sogou Input Method aims to evolve into a "super entrance" for AI, allowing users to interact with AI agents through the input method, which is seen as a key direction for future development [10][11]. - Baidu Input Method also seeks to position itself as a reliable smart entry point for user expression, leveraging AI to enhance communication and collaboration [11][12]. Commercialization Challenges - Input methods face a "high traffic, low value" monetization dilemma, struggling to effectively convert user engagement into revenue [16][17]. - Privacy concerns have led to increased scrutiny and regulatory challenges, prompting companies to adapt their data collection practices to ensure user safety and compliance [18][19]. Technological Limitations - Despite advancements in AI algorithms, challenges remain in accurately understanding user intent, particularly in word prediction, due to the variability in user language habits and the quality of training data [13][19]. - The complexity of multi-modal interactions, where users input data through various means, requires sophisticated algorithms and technology to ensure seamless processing [19].

二、政策与技术环境 一、背景与核心诉求 随着人工智能技术飞速发展，AI 基础设施成为产业智能化变革的核心支撑，在国家 "十四五" 数字经济发展规划与 "东数西算" 工程推动下，我国算力中 心建设呈现政策驱动显著、智能化需求爆发、应用场景深化三大特征。但与此同时，AI 基础设施面临合规要求收紧、云平台漏洞频发、大模型新型攻击 涌现等多重安全挑战，构建高效、安全、可靠的 AI 基础设施成为行业刚需。 （一）整体安全架构 百度构建了以合规为纲、技术为骨、管理为翼的全方位多层级安全防护体系，涵盖合规与标准规范、核心安全域分层防护、管理与运行体系三大维度，形 成 "边界 - 平台 - 租户 - 密码 - 模型 - 运营" 六层联动防护，适配 AI 基础设施特性。 （二）关键安全域防护 模型应用安全：从语料安全清洗、输入输出安全管控、数据安全保护、安全评测能力建设四方面入手，通过 prompt 审核、回复干预、数据加密、蓝军评 测等手段，覆盖大模型训练、部署、推理全生命周期。 云平台安全：以 "治理 - 防护 - 合规 - 运营" 为主线，涵盖安全治理（漏洞、基线）、计算 / 存储 / 网络安全、等保与密评合规、物理安全 ...

Core Insights - Baidu's Apollo Go has received the first fully autonomous driving test license in Dubai, allowing operation of self-driving vehicles without safety drivers on designated public roads [1] - The company plans to launch commercial autonomous ride-hailing services in Dubai within the first quarter of this year, supported by a 2000 square meter autonomous operations center in the city center [1] - The Dubai Roads and Transport Authority (RTA) aims to evaluate system safety, reliability, and user experience through this testing, paving the way for large-scale commercialization [1] Company Developments - Baidu's Apollo Go is set to expand its fleet to over 1000 vehicles in collaboration with RTA [1] - The establishment of the operations center integrates smart roads, charging, maintenance, and scheduling systems [1] Industry Implications - This initiative marks a significant institutional breakthrough for Chinese autonomous driving technology in the Middle East market [1] - The move reflects the UAE's ongoing policy of openness in the field of smart transportation [1]

在此前关于AI产业链的讨论中，海豚君分别从主要聚焦北美市场，阐述了产业链中上游各方"剪不断、理不乱"的利益博弈关系。 而在去年才可以算得上是正式破圈的中国AI，因为消费土壤的因素，在DeepSeek打破算力成本僵局后，默契地选择了以应用落地为竞争点的发展路线， 这也意味着投入性价比是绕不开的硬性指标。 这相比北美的烧钱大户，中国大厂在前置投入上保持了一丝理性。但同时也因为更聚焦终端落地，因此厂商们争相将成熟的、不成熟的AI产品推广曝光 给更多的用户，硬是推着AI的用户渗透在一年之间大大提速。 那么海豚君也从应用端视角，和大家讨论：过去一年的AI发展到什么进度了？中国大厂们之间的战况如何？本篇为行业篇，侧重行业变化和竞争格局， 后续将逐一开展公司篇的讨论。 以下是详细分析 01 AI消费互联网："Byte"靠边，"Token"才是未来 在展开中国AI消费互联网争霸赛之前，海豚君先带大家拎清楚AI时代一场"悄然发生"生产资料大重置。 在AI到来之前的消费互联网，用户使用互联网服务，比如说通过APP和网页，使用搜索、短视频、购物等，除了自己要买个流量/宽带包，基本都是免费 的。 而网页和APP的开发者要提供这些服务 ...

Core Insights - Baidu Baike is launching a comprehensive upgrade centered around AI, introducing "Dynamic Baike" and "Baike AI Knowledge Graph" to revolutionize knowledge dissemination as it approaches its 20th anniversary [1][3] Group 1: Product Innovations - "Dynamic Baike" utilizes generative UI technology to transform static knowledge into interactive, visual content, allowing users to engage with information in a more immersive way [4][6] - The "Baike AI Knowledge Graph" addresses the fragmentation of traditional entries by creating a structured knowledge network that enhances user exploration through multi-dimensional connections [8][10] Group 2: Collaboration and Content Creation - Baidu Baike's "Star Plan" collaborates with over 100,000 experts and professional creators from leading institutions to build a vast repository of over 1 million specialized knowledge entries [3][7] - A nationwide content creation channel will be opened by the end of January, enabling users to contribute to the "Dynamic Baike" platform through interactive dialogue [6] Group 3: Quality Assurance and Authority - Baidu Baike has established a robust content assurance system over its 20 years, ensuring the reliability of its 30 million entries through strict quality control mechanisms [11][12] - The platform emphasizes the importance of verified content in the age of AI-generated misinformation, positioning itself as a critical resource for accurate knowledge [11][12] Group 4: Challenges and Opportunities - Baidu Baike faces challenges from shifting user preferences towards AI direct answers and the proliferation of low-quality AI-generated content [12] - However, the company sees opportunities in reinforcing its role as a reliable knowledge infrastructure, leveraging its extensive database for fact-checking and enhancing user experience through dynamic content [12]

进入 2026，硅谷最会押趋势的那批人，开始强调一个更底层的逻辑：AI 不是先变强，而是先变便宜。 1月7日，a16z 创始人 Marc Andreessen（马克·安德森）在自家播客上做了一场访谈，核心观点是： AI 是他见过最大的技术变革，但关键不在于模型能力的突破，而在于智能本身正在从奢侈 品变成日用品。 调用一次 AI 的成本，正在以惊人的速度往下掉。不是降了一点，是断崖式暴跌。 Marc 同时强调：现在还早。成本已经降下来了，但大部分机会还没被创业者发现。 他这场访谈不谈技术前景，也不谈市场泡沫，而是聚焦一个更具体的问题：如果智能像水电一样便宜且 随处可得，商业规则会怎么变？ 变化体现在四个方面：成本结构、技术路径、定价模式、竞争格局。 现在的 AI，不是谁更强，而是谁先把便宜智能变成标准流程。 第一节｜崩的是价格，起飞的是收入 Marc Andreessen 首先指出：AI 的智能成本正在暴跌 他说： "AI 的单位成本，下降速度比摩尔定律还快。" 模型越训练越强，但每次调用 AI 所需的成本，反而越来越低。不是降一点点，而是断崖式往下掉。 他特别提到：过去一年，大模型的 token 成本正在快速 ...

Group 1 - The core point of the news highlights the short-selling ratios and amounts for various companies, indicating significant market activity and investor sentiment towards these stocks [1][2][3] Group 2 - The top three companies by short-selling ratio are Anta Sports-R (82020) and Li Ning-R (82331) at 100.00%, and Tencent Holdings-R (80700) at 90.68% [1][2] - The highest short-selling amounts are recorded for Alibaba-W (09988) at 3.199 billion, Meituan-W (03690) at 1.835 billion, and Xiaomi Group-W (01810) at 1.518 billion [1][3] - Tencent Holdings-R (80700) has the highest deviation value at 45.50%, followed by Beike-W (02423) at 40.51% and China National Offshore Oil-R (80883) at 40.26% [1][3]