2025年08月25日 证券研究报告 | 人工智能行业专题（11） 探究模型能力与应用的进展和边界 行业研究 · 行业专题 互联网 · 互联网II 投资评级：优于大市（维持） 证券分析师：张伦可 证券分析师：陈淑媛 证券分析师：刘子谭 证券分析师：张昊晨 0755-81982651 021-60375431 liuzitan@guosen.com.cn zhanghaochen1@guosen.com.cn zhanglunke@guosen.com.cn chenshuyuan@guosen.com.cn S0980525060001 S0980525010001 S0980521120004 S0980524030003 请务必阅读正文之后的免责声明及其项下所有内容 报告摘要 Ø 风险提示：宏观经济波动风险、广告增长不及预期风险、行业竞争加剧风险、AI技术进展不及预期风险等。 请务必阅读正文之后的免责声明及其项下所有内容 2 Ø 本篇报告主要针对海内外模型发展、探究模型能力与应用的进展和边界。我们认为当前海外模型呈现差异化发展，企业调用考虑性价比。当前 OpenAI在技术路径上相对领先，聚焦强化推理与专业 ...

Core Viewpoint - The recent upgrade of DeepSeek to version 3.1 has shown significant improvements in context length and user interaction, while also merging features from previous models to reduce deployment costs [1][11][12]. Group 1: Model Improvements - DeepSeek V3.1 now supports a context length of 128K, enhancing its ability to handle longer texts [4]. - The model's parameter count increased slightly from 671 billion to 685 billion, but the user experience has improved noticeably [5]. - The model's programming capabilities have been highlighted, achieving a score of 71.6% in multi-language programming tests, outperforming Claude 4 Opus [7]. Group 2: Economic Efficiency - The merger of V3 and R1 models allows for reduced deployment costs, requiring only 60 GPUs instead of the previous 120 [12]. - Developers noted that the performance could improve by 3-4 times with the new model due to increased cache size [12]. - The open-source release of DeepSeek V3.1-Base on Huggingface indicates a move towards greater accessibility and collaboration in the AI community [13]. Group 3: Market Context - The AI industry is closely watching the developments of DeepSeek, especially in light of the absence of the anticipated R2 model [19]. - Competitors like OpenAI, Google, and Alibaba have released new models, using R1 as a benchmark for their advancements [1][15]. - The market is eager for DeepSeek's next steps, particularly regarding the potential release of a multi-modal model following the V3.1 update [23].

Core Viewpoint - OpenAI's founder Sam Altman believes that the U.S. is underestimating the threat posed by China's next-generation artificial intelligence, and that chip regulations alone are not an effective solution [1][3] Group 1: AI Competition - Altman stated that China can develop faster in reasoning capabilities and has strengths in research and product development [3] - The AI competition between the U.S. and China is deeply intertwined, going beyond simple rankings [3] Group 2: OpenAI's Strategic Shift - OpenAI recently released its first open-weight models, gpt-oss-120b and gpt-oss-20b, marking a significant strategic shift from its long-standing closed-source approach [3] - The decision to release open-weight models was influenced by competition from Chinese models, such as DeepSeek and Kimi K2 [3] - Altman emphasized that if OpenAI did not act, Chinese open-source models would gain widespread adoption, making this a significant factor in their decision [3]

Group 1 - The Chinese AI industry is rapidly rising, with the release of the Kimi K2 model being a significant milestone, indicating a new wave of technological advancement driven by AI [1] - The Hong Kong stock market is attracting global capital, with a notable inflow of over 215 billion HKD in May, June, and July, averaging 5.1 billion HKD daily [1] - The launch of the Huaxia National Index Hong Kong Stock Connect Technology ETF on August 18 provides investors with a convenient tool to invest in leading Hong Kong tech companies [2] Group 2 - The National Index Hong Kong Stock Connect Technology Index has gained popularity, with its ETF product size increasing from 7.7 billion to 27.8 billion RMB, a growth of 259% [3] - The index selects 30 large-cap, high R&D investment, and fast-growing tech stocks, ensuring a focus on innovation and market potential [3] - The index's sector distribution is balanced, with significant allocations in electronics (23%), media (22%), and pharmaceuticals (15%), highlighting its focus on innovative drug sectors [4] Group 3 - The National Index Hong Kong Stock Connect Technology Index has shown impressive long-term performance, with a cumulative return of 159.9% since 2017, outperforming other indices [4] - The index's valuation metrics indicate a favorable price-performance ratio, with a P/E ratio of 24.46 and a P/S ratio of 2.92, suggesting it is undervalued compared to global tech indices [4] - The leading position of Huaxia Fund in the ETF market is underscored by its management of over 720 billion RMB in equity ETFs, reflecting its strong investment capabilities [5] Group 4 - The Hong Kong tech sector is experiencing a systemic valuation reset, driven by increased AI computing power supply and a rational market return in the delivery industry [6] - The issuance of the Hong Kong Stock Connect Technology ETF offers investors an excellent opportunity to capitalize on the global revaluation of Chinese tech assets [6]

Summary of Conference Call Records Industry Overview - The conference call discusses the AI computing demand in the domestic market and the capital expenditure (CAPEX) trends of overseas cloud service providers (CSPs) [1][2][3]. Key Points on Overseas CSPs - Total capital expenditure of overseas CSPs has reached $350 billion, with a healthy CAPEX to net cash flow ratio of around 60% for all but Amazon, which has higher costs due to logistics investments [2]. - Microsoft and Google have shown significant growth in cloud and AI revenues, alleviating KPI pressures [2]. - Microsoft Azure's revenue growth is significantly driven by AI, contributing 16 percentage points to its growth [5]. - Google has increased its CAPEX by $10 billion for AI chip production, with its search advertising and cloud businesses growing by 11.7% and 31.7% year-over-year, respectively [2]. - Meta has financed $29 billion for AI data center projects, with a CAPEX to net cash flow ratio also around 60%, despite concerns over cash flow due to losses in its metaverse business [2]. AI Profitability Models - The profitability model for overseas CSPs in AI is gradually forming, with a focus on cash flow from cloud services and enhancing traditional business efficiency through AI [5]. - Meta's AI recommendation models have improved ad conversion rates by 3%-5% and user engagement by 5%-6% [5]. - The remaining performance obligations (RPO) for a typical CSP reached $368 billion in 2025, indicating a 37% year-over-year growth, locking in future revenues [5]. AI Model Competition and User Retention - The overall user stickiness of large models is weak, but can be temporarily improved through product line expansion and application optimization [6]. - Deepsec's R1 model held a 50% market share on the POE platform in February 2025 but dropped to 12.2% three months later due to intense competition [7]. - Different large models exhibit unique advantages in specific applications, such as Kimi K2 for Chinese long text processing and GPT-5 for complex reasoning [9]. Domestic AI Computing Demand - Domestic AI computing demand is robust, with a requirement for approximately 1.5 million A700 graphics cards for training and inference [3][12]. - The demand for AI computing is growing faster than chip supply, resulting in a 1.39 times gap, indicating a continued tight supply in the coming years [3][16]. - The total estimated demand for AI computing in the country is around 1.5 million A700 cards, equating to the overall training and inference needs [15]. Video Inference and Overall Demand - Video inference calculations indicate that approximately 100,000 A700 cards are needed for video processing, contributing to a total demand of about 250,000 A700 cards when combined with training needs [13][12]. - The overall AI demand is projected to be very strong, with significant capital expenditure implications [13]. Conclusion - The conference call highlights the growing importance of AI in both domestic and international markets, with CSPs adapting their business models to leverage AI for revenue growth while facing competitive pressures and supply constraints in computing resources [1][2][3][5][16].

Core Viewpoint - The global large model industry continues to develop rapidly post "DeepSeek innovation heat," with an acceleration in model iterations and an increase in computing power demand driven by token consumption [2][7][25]. Group 1: Global Model Updates and Computing Demand - In Q2 2025, major model companies like Google and OpenAI released significant updates, including OpenAI's GPT-5, which improved efficiency and reduced API costs, thus increasing computing power demand [3][13][25]. - The release of GPT-5 marked a shift towards efficiency, with a notable reduction in token consumption and a context window expansion to 400K tokens, enhancing application capabilities and driving further demand for computing resources [18][22][25]. - The North American model updates have created a preliminary closed loop in computing power demand, with companies like Google and Anthropic seeing rapid increases in token consumption [3][30]. Group 2: Domestic Model Development and Market Dynamics - Domestic companies, while still trailing behind in model capabilities, have made significant strides since 2025, with firms like ByteDance and Kimi releasing updated models that have increased computing power consumption [4][36]. - The domestic AI chip industry is evolving from single-chip solutions to system-level designs, supporting the iteration and deployment of large models [4][43]. - The anticipated updates from open-source models like DeepSeek in Q3 2025 could reignite investment sentiment in the domestic AI industry [4][36]. Group 3: Token Consumption Trends - Token consumption has surged globally, with major players like Google, Microsoft, and ByteDance experiencing significant increases in token usage since 2025 [27][30]. - Google's AI Overview feature has been a key driver of its token consumption growth, leveraging its vast user base to generate high-frequency AI summaries [30][31]. - The current market dynamics reflect a balance between free access to AI technologies and the monetization of high-value applications, with paid products showing a clear differentiation in performance and reliability [34][35]. Group 4: Future Outlook and Investment Opportunities - The ongoing model updates and the increasing efficiency of token usage are expected to drive sustained growth in computing power demand, with both cloud and edge computing becoming critical [22][24][36]. - The competitive landscape suggests that companies with robust financial backing, like Google and Meta, will continue to push for model updates, further enhancing computing demand [26][30]. - The domestic AI industry is poised for growth as local firms enhance their capabilities and seek to capture market share in the evolving landscape of AI applications [4][43].

Core Viewpoint - The article discusses the tension between the grand narrative of AI and the practical challenges faced by product managers in implementing AI solutions, highlighting the gap between theoretical concepts and real-world applications [1][2][9]. Group 1: AI Product Development Challenges - Product managers are overwhelmed by the rapid advancements in AI technologies, such as GPT-5 and Kimi K2, while struggling to deliver a successful AI-native product that meets user expectations [1][2]. - There is a significant divide between those discussing the ultimate forms of AGI and those working with unstable model APIs, seeking product-market fit (PMF) [2][3]. - The current AI wave is likened to a "gold rush," where not everyone will find success, and many may face challenges or be eliminated in the process [3]. Group 2: Upcoming Global Product Manager Conference - The Global Product Manager Conference scheduled for August 15-16 aims to address these challenges by bringing together industry leaders to share insights and experiences [2][4]. - Attendees will hear firsthand accounts from pioneers in the AI field, discussing the pitfalls and lessons learned in transforming AI concepts into viable products [5][6]. - The event will feature a live broadcast for those unable to attend in person, allowing broader participation and engagement with the discussions [2][11]. Group 3: Evolving Role of Product Managers - The skills traditionally relied upon by product managers, such as prototyping and documentation, are becoming less relevant due to the rapid evolution of AI technologies [9]. - Future product managers will need to adopt new roles, acting as strategists, directors, and psychologists to navigate the complexities of AI integration and user needs [9][10]. - The article emphasizes the importance of collaboration and networking in this uncertain "great maritime era" of AI development [12].

7月12日，中国某AI公司发布的Kimi K2开源模型成为关注焦点，在国际上获得了"又一个DeepSeek时 刻"的评价。随后，8月7日，OpenAI发布GPT-5，引发了海内外的广泛讨论。AI大模型接连更新，对投 资而言意味着什么？（资料参考：财通证券研究《财通计算机·中美AI百花齐放，开启AI新时代》， 2025.7.21） 新发布的大模型有哪些核心优势？ 在技术层面，Kimi K2的总参数达1万亿（1T），是当前大模型参数量的天花板。从多个基准测试成绩 来看，此次Kimi K2超过了DeepSeek-V3-0324、Qwen3-235B-A22B等开源模型，成为开源模型新SOTA （当前最高水平）。（资料参考：青橙财经《抢先DeepSeek R2，开源万亿参数Kimi K2：月之暗面生死 突围》，2025.7.21：机器之心《深夜开源首个万亿模型K2，压力给到OpenAI，Kimi时刻要来了？》， 2025.7.12） 2．控价开源：高性价比策略吸引用户 价格上，Kimi K2延续了"高性价比"策略：每百万输入tokens收费4元，每百万输出tokens收费16元。综 合来看，API调用成本与DeepS ...

Core Viewpoint - The investment landscape for AI large models has become cautious, with investors preferring to bet on the top two players, leading to a shrinking space for mid-tier and lower players [2][45]. Group 1: Product Developments - OpenAI released GPT-5, integrating large language models with reasoning models, significantly reducing factual error rates compared to GPT-4o [4][5]. - Kimi K2, a trillion-parameter model from the domestic AI company "月之暗面," gained significant attention with 3.6 billion website visits and over 100,000 downloads within 48 hours of its launch [7]. - Despite initial hype, Kimi K2 failed to replicate the success of DeepSeek, with its monthly active users (MAU) dropping from 21.01 million in December to 14.08 million in May, ranking ninth among domestic AI applications [10][9]. Group 2: Marketing and User Acquisition - Kimi's rapid user growth was attributed more to aggressive marketing strategies rather than technological superiority, with over 100 million yuan spent on advertising during the 2024 Spring Festival [17]. - The high user acquisition cost (CPA) of 30 yuan per user reflects the heavy investment in marketing, which has led to a temporary spike in attention and downloads [17]. Group 3: Competitive Landscape - Kimi's long-text processing capabilities, once seen as a competitive advantage, have been diminished as other models like 通义 and 豆包 have surpassed its performance [22][25]. - The competition has shifted towards reasoning ability, interaction experience, and multimodal capabilities, with Kimi lagging in these areas [28][30]. Group 4: Challenges and Limitations - Kimi faces significant challenges due to insufficient data access, relying heavily on public data and limited partnerships, unlike competitors who leverage their ecosystems for continuous model optimization [35][36]. - The company is constrained by a lack of high-performance computing resources, particularly after the U.S. restrictions on advanced chips, which hampers its ability to compete with top global models [37][38]. - Capital constraints have emerged as a new pressure, with Kimi's valuation peaking at $3.3 billion but facing a lack of new funding since August 2022, leading to a potential cycle of reduced innovation and user loss [41][42].

Core Insights - Moonshot AI, based in Beijing, is gaining attention for its open-source model Kimi K2, which ranked fifth globally upon its launch in July 2023 [1] - The company's mission is to explore the limits of intelligence and make AI universally accessible [1] Company Overview - Founded in April 2023 by a team with extensive experience in natural language processing (NLP), Moonshot AI aims to discover transformative possibilities in artificial intelligence [1] - The company has approximately 300 employees, with a significant portion being young talent from the '90s generation [2] Product Development - Kimi K2, a trillion-parameter model, has a unique capability to handle long texts, supporting up to 200,000 Chinese characters [2][5] - The Kimi intelligent assistant was launched in October 2023, followed by several product releases, including Kimi browser assistant and Kimi-Researcher [2] Technical Innovations - Kimi K2's architecture allows for complex tasks at a lower cost, with only 32 billion active parameters [3] - The model has excelled in various benchmarks, particularly in programming, tool usage, and mathematical reasoning [6] User Engagement - Kimi K2's long-text capability has led to a significant increase in user adoption, with user numbers growing from hundreds of thousands to tens of millions in 2024 [5] - The model is designed to be user-friendly, allowing non-programmers to utilize its capabilities effectively [7] Future Aspirations - Moonshot AI aims to create a general-purpose AI that surpasses human intelligence, focusing on developing versatile skills that can enhance each other [8] - The company emphasizes the importance of building a strong foundational model before releasing products, ensuring robust performance and capabilities [8]