生成引擎优化（GEO）是从传统搜索引擎优化（SEO）演变而来的新兴技术， 旨在通过 AI 工具提高内容在搜索引擎和社交媒体平台上的曝光率。它的核心逻 辑包括了解目标客户群体、精准营销和内容分发。百度在 GEO 方面的应用已经 占据了其广告搜索业务约 40%的份额，每年营收中的 2/3 来自广告搜索业务。 百度的 GEO 业务布局可以分为三大类：技术供给链、内容服务生态和营销解决 方案。在技术供给链方面，百度提供标准化的 SDK 包和平台工具，开发者可以 利用这些工具进行二次开发和定制。在内容服务生态方面，百度通过全国 20 多个数据众包基地获取大量数据标注素材，为 AI 训练提供支持。在营销解决方 案方面，百度提供端到端的策略服务，通过词包类服务帮助客户实现精准投放。 百度如何通过技术工具链实现 GEO？ 百度通过技术工具链实现 GEO 主要依赖于 RAG 优化、大模型基模和向量数据 库检索等技术。具体来说，开发者可以使用百度提供的 SDK 包或 API 接口，将 需要优化的内容导入系统，并选择投放的平台，如 Deepseek、豆包、百度文 心等。系统会根据用户需求生成策略，通过几分钟到几十分钟的时间完成优化 ...

关于港股后市 本文转载自腾讯自选股，智通财经编辑：李佛。 中信证券认为，港股方面，受益于内部的"十五五"催化以及外部主要经济体的"财政+货币"双宽松，港 股2026年有望迎来第二轮估值修复以及业绩进一步复苏的行情，建议关注科技、大医疗、资源品、必选 消费、造纸和航空板块。 招商证券认为，市场经历拔估值后，叠加宏观内需约束限制指数上行天花板，将进入由质量驱动的结构 性分化阶段。该机构建议双线并行，恒科反弹+增强哑铃：聚焦超跌反弹和成长，优先配置估值回归合 理的恒生科技指数(AI互联网龙头等)。哑铃增强：进攻端：有色金属(商品增强)+保险板块(权益Beta增 强)。防守端：高股息资产(固收增强)作为"压舱石"。 1月12日，恒生指数高开0.55%，恒生科技指数涨0.88%。盘面上，科网股活跃，美团、百度集团均涨超 2%;锂矿股强势，赣锋锂业、天齐锂业均涨超4%;贵金属板块再度走强，紫金矿业涨近3%，中国铝业涨 超2%。 浙商国际认为，港股市场基本面仍偏弱，资金面环境有所回落，政策面重点关注新质生产力和扩大内 需，情绪面短期回暖。当下港股市场周月线级别趋势仍位于右侧区间，对于后续走势，即使短期行情有 波折，我们仍不 ...

| 股票名称 | 沽空金额↓ | 沽空比率 | | 偏离值 | | --- | --- | --- | --- | --- | | 美团-W(03690) | 15.54 亿元 | 21.96% | 4.92% | | | 阿里巴巴-W(09988) | 14.40 亿元 | 8.47% | -6.76% | | | 腾讯控股(00700) | 12.53 亿元 | 11.49% | 0.80% | | | 中国平安(02318) | 10.67 亿元 | 31.14% | 3.29% | | | 百度集团-SW(09888) | 6.38 亿元 | 31.55% | 3.08% | | | 联想集团(00992) | 6.08 亿元 | 40.42% | 11.00% | | | 中国石油化工股份 (00386) | 5.62 亿元 | 27.88% | 2.19% | | | 快手-W(01024) | 5.46 亿元 | 18.47% | 3.10% | | | 友邦保险(01299) | 5.23 亿元 | 32.73% | 11.86% | | | 建设银行(00939) | 5.08 亿元 | 3 ...

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 ...