证券研究报告 总量研究 / 策略研究 2026 年 03 月 30 日 寻找中国的 HALO 资产 ——把握 AI 时代中国的 HALO 资产配置机遇 投资要点： 证券分析师 前言：2026 年 2-3 月，HALO 资产获得投资者追捧，全球掀起了 HALO 资产热。 聚焦 HALO：AI 引发 HALO 热，地缘助推 HALO 热，继续看好 HALO 资产。 HALO 是"Heavy Assets, Low Obsolescence"的缩写，即"重资产、低淘汰"。HALO 概念的 兴起，是 AI 技术革命在资本市场中引发的深刻投资逻辑迁移——当"轻资产、高增长"类科技 企业的吸引力下降时，具有高壁垒有形资本、难以被人工智能替代的传统重资产企业，重新 获得全球投资者的广泛关注。 与过往多年"轻资产、高增长"资产受欢迎的格局迥异，HALO 资产成为市场关注焦点，其 原因有三： 第一 ，美国科技巨头资本开支增速放缓，投资者降低关注； 第二 ，AI 在技术突飞 后应用爆发，各行各业面临前所未有的变革，一些"轻资产、高增长"行业甚至会被 AI 颠覆 或替代，这给相关行业带来了焦虑，投资者对相关行业的关注度进一步降低； ...

Investment Rating - The report maintains a "Buy" rating for the computer industry [2] Core Insights - OpenClaw presents extraordinary opportunities for Chinese large model manufacturers, as the commercialization of these models is not yet fully closed compared to overseas counterparts [4][11] - OpenClaw is transforming the landscape by standardizing and quantifying API calls, allowing for a clear conversion of interactions into pay-per-use revenue streams, thus enabling Chinese large model manufacturers to achieve a commercial closure in the C-end market [5][15] - The C-end market in China has not yet formed a subscription-based revenue model, unlike the mature overseas market represented by ChatGPT, which has over 900 million weekly active users and more than 50 million individual subscribers [7][14] Market Review - The computer industry index fell by 3.44% this week, ranking 30th among the primary industries [21] - The overall price-to-earnings (PE) ratio for the computer industry was reported at 50.4 as of March 27 [24] Investment Recommendations - The report suggests focusing on ecological domestic large models that have high consumption ratios within the OpenClaw ecosystem, as their API call volumes and token revenues are expected to enter a high growth phase [26] - Attention should also be given to leading cloud service providers that offer one-click deployment solutions and control application distribution entry points [26] - The demand for inference-side computing power is expected to surge, benefiting domestic AI chip manufacturers, servers, and related hardware suppliers [26]

Core Viewpoint - The article emphasizes the rapid evolution of AI applications and the importance of selecting reliable AI tools through authentic evaluations, highlighting the launch of the 36Kr AI Evaluation Mini Program as a solution to navigate the crowded AI landscape [6][13]. Group 1: AI Application Landscape - The article notes a surge in AI applications, with over 400 AI tools already cataloged in the 36Kr AI Evaluation platform, covering various categories such as office work, programming, design, and daily life [13][16]. - It mentions that the domestic large model count has surpassed 1,500, indicating a significant growth in AI capabilities and options available to users [16]. Group 2: Features of 36Kr AI Evaluation - The platform offers a product navigation feature that allows users to filter AI tools by categories like "text processing," "video processing," and "educational assistance," making it easier to find suitable applications [19]. - The AI product review leaderboard aggregates popular reviews, helping users identify trending and lesser-known but effective tools [20]. Group 3: Community Engagement - The platform encourages user-generated content, allowing individuals to share their experiences and evaluations of various AI tools, thus fostering a community of shared knowledge [26][29]. - Users can explore a "Discovery Square" to read others' evaluation notes and join interest-based circles for targeted learning, enhancing the overall user experience [25].

Cursor admits its new coding model was built on top of Moonshot AI’s Kimi https://t.co/Fdorm24WKV ...

Core Insights - Tesla and xAI founder Elon Musk have shown support for Chinese domestic large models, which has attracted attention in the industry [1][3] - The global programming tool Cursor released its self-developed coding model Composer 2, which surpassed Claude Opus 4.6 in evaluations and emphasizes cost-effectiveness [1] - Composer 2 is based on the Kimi K2.5 model, which Musk acknowledged on social media [1] Group 1 - The Kimi team expressed gratitude using the Chinese phrase "Thank you for being you," showcasing a blend of technical confidence and warmth [3] - On March 16, Kimi released a technical report titled "Attention Residuals," which restructured the residual connection mechanism of large models, achieving a 1.25 times improvement in training efficiency on a 48 billion parameter model, with scientific reasoning and mathematical performance increasing by 7.5% and 3.6% respectively [3] - Musk praised Kimi's work on social media, highlighting the impressive nature of their achievements [3] Group 2 - On March 2, Alibaba's Qwen officially open-sourced four small-sized models: Qwen3.5-0.8B, 2B, 4B, and 9B, which Musk commented on, noting the impressive intelligence density [3] - ByteDance's new video generation model Seedance 2.0 began internal testing on February 12, addressing industry pain points such as low usability and character detail drift, capable of generating 60 seconds of 2K broadcast-quality video [3] - Musk expressed amazement at the rapid advancements in AI technology, stating "It's happening fast" in response to developments in Seedance 2.0 [3]

Core Insights - The financial industry is experiencing a significant shift towards AI-driven efficiency, with a reported replacement rate of 94% for financial positions, indicating a vast potential for automation and cost reduction [3][8][9] - The emergence of AI tools, such as "OpenClaw," is enabling firms to enhance research efficiency and reduce reliance on human researchers, leading to a transformation in the investment landscape [10][11][12] Industry Trends - Financial professionals are under constant pressure to improve performance and adapt to new technologies, with AI tools being integrated into various aspects of investment management [7][8] - The cost of human labor in the financial sector is high, with quantitative researchers earning between 800,000 to 1,500,000 yuan annually, while AI can potentially save millions in operational costs [9][10] - Private equity firms are increasingly adopting AI technologies to streamline operations, with some firms reporting that AI can outperform human researchers in efficiency [11][12] AI Integration - The integration of AI in investment research is seen as a way to eliminate inefficiencies and reduce the need for a large workforce, with AI agents capable of working continuously without the costs associated with human employees [9][10] - The development of AI tools is leading to a new paradigm where traditional roles in investment management may become obsolete, as firms seek to optimize their operations [12][14] - The financial industry is witnessing a shift where AI tools are not just augmenting human capabilities but are beginning to replace them in certain functions, raising questions about the future roles of human researchers and fund managers [16][18] Challenges and Considerations - Despite the advantages of AI, there are concerns regarding the reliability and safety of these technologies, particularly in quantitative finance, where randomness and uncertainty can pose risks [13][14] - The financial sector is grappling with the implications of AI on job roles, with some professionals questioning the necessity of human researchers if AI can fulfill their functions [16][18] - The rapid advancement of AI is creating a sense of urgency within the industry, as firms strive to keep pace with technological developments and avoid being left behind [17][18]

Core Viewpoint - The article discusses the significant impact of AI on the financial industry, particularly in asset management, highlighting the potential for job displacement and the shift towards automation in research and operations [6][11]. Group 1: AI Impact on Employment - A recent report indicates that the job replacement rate in finance could reach 94%, with the current actual rate at 28%, suggesting a vast potential for future displacement [6]. - Financial professionals are encouraged to consider alternative employment opportunities, such as dishwashing or plumbing, as AI continues to evolve [6]. Group 2: Cost Efficiency and AI Adoption - In a competitive environment with high operational costs and diminishing alpha, private equity firms are focusing on optimizing human efficiency [12]. - Salaries for quantitative researchers typically range from 800,000 to 1,500,000 yuan, with significant bonuses for successful recommendations, indicating high costs associated with human labor [13]. - The adoption of AI in research could save millions in costs, as AI can work continuously without the expenses associated with human employees [13]. Group 3: AI Tools and Research Efficiency - The emergence of tools like OpenClaw is seen as a way to enhance research efficiency, with claims that it can significantly increase productivity [14][21]. - Private equity firms are increasingly integrating AI into their operations, with some firms reporting that AI agents can outperform human researchers in terms of efficiency [17][21]. Group 4: Challenges and Limitations of AI - Despite the enthusiasm for AI, some experts express skepticism about its effectiveness in serious quantitative environments, suggesting that tools like OpenClaw may not meet the rigorous demands of quantitative investment [20]. - Concerns are raised about the randomness and safety of AI tools, which could introduce significant uncertainty into quantitative systems [20]. Group 5: Future of Human Roles in Finance - As AI takes over more tasks, the role of human researchers may shift, with expectations for them to focus on specific tasks rather than broader market analysis [23]. - The article suggests that if AI can fulfill the roles of researchers and fund managers, the need for human oversight may diminish, raising questions about the future of the industry [27]. Group 6: Industry Sentiment and Adaptation - The financial industry is experiencing heightened anxiety due to rapid advancements in AI, leading to a culture of continuous learning and adaptation [26]. - The article emphasizes the importance of defining the roles of humans and AI in the financial sector, suggesting that collaboration rather than competition may be the key to future success [28].

Core Insights - Tesla and xAI founder Elon Musk have been actively engaging with China's AI sector, praising domestic models such as Kimi, ByteDance's Seedance 2.0, and Alibaba's Qwen3.5, indicating a significant cross-ocean technology interaction [1][2] Group 1: Kimi Model - Kimi released a technical report titled "Attention Residuals," which restructured the residual connection mechanism of large models, achieving a 1.25 times improvement in training efficiency on a 48 billion parameter model, with scientific reasoning and mathematical performance increasing by 7.5% and 3.6% respectively, marking a significant signal for "Deep Learning 2.0" [1] - Musk praised Kimi's work on social media, highlighting its impressive nature, to which Kimi humorously responded [1] Group 2: Alibaba's Qwen3.5 - Alibaba's Qwen3.5 series models were recognized for their "extreme intelligence density," with the official release of four small-sized models (0.8B, 2B, 4B, 9B) on March 2, designed to meet diverse needs from edge devices to lightweight servers, breaking the stereotype that smaller models have weaker capabilities [1][2] - The 9B version of Qwen3.5 is reported to perform comparably to models with hundreds of billions of parameters, while the 0.8B and 2B versions can run smoothly on mobile and IoT edge devices [2] - Musk commented on the impressive intelligence density of Qwen, further emphasizing its significance [2] Group 3: ByteDance's Seedance 2.0 - ByteDance's Seedance 2.0, a next-generation video generation model, began internal testing on February 12, featuring a unified multi-modal audio-video generation architecture that supports text, image, audio, and video inputs, addressing industry pain points such as low usability and character detail drift [2] - The model can generate up to 60 seconds of 2K broadcast-quality video, showcasing its advanced capabilities [2] - Musk expressed astonishment at the rapid advancements in this area, indicating a fast-paced evolution in AI technology [2] Group 4: Musk's Predictions on China's AI - Musk has predicted that China's AI computing power will surpass that of other regions, citing stable and inexpensive electricity, large-scale infrastructure, and an efficient engineering workforce as core advantages for China's AI development [2]

Core Insights - The large model landscape has become increasingly crowded with numerous models emerging rapidly, making it difficult to understand their architectures and innovations [2][3] - A significant gap exists in the availability of a clear visual representation of these models, despite the abundance of options [2] Summary by Sections - **Introduction to the Landscape**: The article highlights the rapid development of large models such as GPT, Llama, and others, noting the challenge in comprehending their diverse architectures [2] - **Creation of the LLM Architecture Gallery**: AI researcher Sebastian Raschka has created an online resource called the "LLM Architecture Gallery," which organizes and visualizes the architectures of mainstream large models [3][6] - **Content of the Gallery**: The gallery serves as a comprehensive directory of various models, ranging from those with millions to trillions of parameters, including notable names like Llama, DeepSeek, and Mistral [7] - **Model Cards**: Each model in the gallery is linked to a dedicated page that provides essential information such as architecture diagrams, key module designs, parameter sizes, and release dates, facilitating quick understanding for researchers [11][14] - **Utility for Researchers**: The gallery acts as a quick reference index for model architectures, allowing users to compare different designs and innovations efficiently, thus aiding in understanding the evolution of technology [14]

Core Viewpoint - The article discusses the emergence of Generative Engine Optimization (GEO) as a new business model in the AI landscape, highlighting its rapid growth and the associated risks of misinformation and manipulation within AI models [5][10][30]. Group 1: GEO Business Model - GEO has seen explosive growth in the past year, with many businesses seeking to influence AI-generated answers to increase product visibility and traffic [6][11]. - The core purpose of GEO is to affect AI-generated responses, ensuring that products or brands appear prominently in the answers provided by AI models [10][12]. - The service providers in the GEO space have surged, with estimates suggesting that there are hundreds of companies now offering GEO services, reflecting a highly competitive market [11][12]. Group 2: Market Dynamics and Challenges - The traditional growth methods in the mobile internet space have plateaued, leading brands to explore GEO as a new avenue for traffic generation [13][28]. - The effectiveness of GEO is often overstated, as it tends to function more like brand advertising rather than direct response advertising, with low conversion rates [28][40]. - The market for GEO services is characterized by high levels of service homogeneity, with pricing ranging from thousands to tens of thousands of yuan based on keyword or question volume [14][28]. Group 3: Technical Aspects of GEO - GEO's operational process involves creating customized content based on client information, which is then distributed across various platforms to influence AI models [14][19]. - The effectiveness of GEO relies on understanding the preferences of different AI models, which can vary significantly, necessitating tailored content strategies [19][21]. - The content produced for GEO must be structured and information-dense to avoid being flagged as promotional material by AI models [21][24]. Group 4: Risks and Ethical Concerns - The practice of "poisoning" AI models with misleading information has been highlighted, where companies manipulate training data to favor their products [30][33]. - The prevalence of low-quality AI-generated content poses a significant challenge, as it can degrade the overall quality of information available to users [40][41]. - As the GEO market matures, there is a growing concern about the sustainability of such practices, with potential regulatory responses anticipated from AI model providers [34][42]. Group 5: Future Outlook - The GEO landscape is expected to evolve as AI platforms begin to implement clearer commercial rules, potentially reducing the gray areas currently exploited by service providers [51][52]. - Companies are encouraged to build a robust online presence and provide high-quality content to improve their visibility in AI-generated responses [48][50]. - The competition for visibility in AI models is likened to a trust game, where companies must engage meaningfully with AI rather than relying on manipulative tactics [47][52].