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 Viewpoints - The restructuring of the industrial ecosystem through anti-involution is beneficial for sustainable profitability in various sectors [2][3] - Despite significant advancements in China's technology sector, market sentiment remains pessimistic, impacting the performance of tech stocks [2][5] - The Chinese manufacturing industry plays a crucial role in the global supply chain, with trade risks currently deemed manageable [2] Industry Insights - The current excessive internal competition in industries is hindering wealth accumulation and income growth, leading to layoffs and reduced innovation investment [3] - Legislative measures against unfair competition are being implemented, marking a shift towards a more regulated market environment [3] - The AI sector is experiencing rapid growth, with significant investments from global tech giants, indicating a strong future demand for AI-related infrastructure and applications [6][12] Investment Opportunities - The focus is on sectors aligned with new productive forces, particularly in technology and pharmaceuticals, which are expected to benefit from economic recovery [5][7] - The pharmaceutical sector is seen as having long-term investment value due to demographic trends and the potential for innovative drug development [7][13] - The technology sector's valuation is becoming attractive, especially in areas like AI and consumer electronics, as the market adjusts to new pricing realities [8][10] Macro Economic Context - The macroeconomic environment remains complex, with ongoing uncertainties in trade and economic fundamentals, leading to a focus on structural opportunities in high-performing sectors [10][15] - The Chinese asset market is gradually showing value, with expectations for improved profitability and market performance in the second half of the year [15] - The automotive industry in China is on the rise, with domestic brands gaining market share and expanding into higher-end segments [17] Sector-Specific Trends - The renewable energy sector is facing challenges due to overcapacity and competition, prompting a need for regulatory changes to foster healthier market conditions [16][18] - Companies with strong growth potential and competitive advantages are being prioritized for investment, particularly in sectors like automotive parts and electronics [19]

Core Insights - DeepSeek's monthly average downloads significantly dropped from 81.13 million in Q1 2025 to 22.59 million in Q2 2025, a decline of 72.2% [2] - The decline is attributed to user diversion to other applications that have integrated DeepSeek's open-source model, with 59.2% of lost users switching to Baidu App and 38.6% to Doubao App [2] - Major companies like Alibaba, ByteDance, and Baidu have launched cheaper competing APIs, further squeezing DeepSeek's market space [2] Company Overview - DeepSeek, developed by Deep Seek (Hangzhou) Technology Co., is an open-source AI product known for its low cost and high performance, with a training cost of only $6 million using 2048 NVIDIA H800 GPUs [3] - Despite the drop in downloads, DeepSeek's open-source strategy has contributed significantly to the industry's development [3] Industry Context - The AI model cost in China is significantly lower than that of international giants, with DeepSeek-R1's inference cost being about one-thirtieth of OpenAI's operational cost [5] - As of April 2024, approximately 305 large models have been launched in China, with 254 of them having over 1 billion parameters [4] Competitive Landscape - Baidu's Wenxin model 4.5 and X1 have been released, with the former outperforming GPT-4.5 in several tests and having an API call price only 1% of GPT-4.5's [5] - The competitive landscape includes various models such as Alibaba's Tongyi Qianwen, ByteDance's Doubao model, and others, each with unique features and pricing strategies [6] Technological Impact - AI technologies represented by DeepSeek are becoming core drivers of industry innovation, enhancing data integration, multi-modal analysis, and complex scenario simulation [7] - The lightweight nature, performance improvements, and rapid cost reductions of large models are accelerating their development and application in new industrialization [9]

Core Insights - The breakthrough in large model technology is driving the development of multimodal and agent technologies, enhancing industry efficiency and accelerating commercialization through policy compliance and capital resonance [1][2][4]. Market Dynamics - By 2025, China's large model technology is expected to experience explosive growth and structural optimization, transitioning from an auxiliary tool to a core productivity driver across various sectors including government, finance, manufacturing, and healthcare [2][4]. - In the first half of 2025, the bidding market for large models reached a record scale of 6.4 billion yuan with 1,810 projects, surpassing the total number of projects in 2024 [4][5]. - Baidu Smart Cloud emerged as the leading bidder with 48 projects and 510 million yuan in bid amounts, followed by iFLYTEK and Volcano Engine [4][5]. Technological Advancements - Significant breakthroughs in multimodal capabilities and agent technologies are fostering a positive cycle of technology, application, and business [7][8]. - The market is shifting focus from infrastructure to practical business applications, with over 50% of projects in the second quarter of 2025 being application-oriented [5][6]. - The integration of large models with industrial software is becoming a mainstream application mode, particularly in manufacturing [11][12]. Policy and Regulatory Framework - A comprehensive policy framework has been established at the national level, focusing on compliance, incentives, and infrastructure to guide the healthy development of the industry [14][15]. - As of June 30, 2025, 439 generative AI services have completed registration, indicating a move towards standardized development [14][15]. Regional Development - Different regions in China are adopting unique development paths for large models, with the Beijing-Tianjin-Hebei region focusing on technological breakthroughs, while the Yangtze River Delta emphasizes scene innovation and ecological cultivation [18][19][20]. Capital Market and Industry Collaboration - The surge in bidding orders for large model vendors is linked to internal innovation and policy support, with significant impacts on stock prices following major project wins [21][23]. - The integration of capital operations through mergers, strategic investments, and industry chain collaboration is accelerating the commercialization of large model technologies [25][26].

Macro and Strategy - The issuance of special bonds for land reserves has reached nearly 170 billion [7] - In the 24th week (June 9-15), the net financing of government bonds was 2,190 billion, and for the 25th week (June 16-22), it is expected to be 2,594 billion, with a cumulative total of 6.8 trillion [7][8] - The total general deficit as of the 24th week is 51 trillion, with a progress rate of 43.4% [7] Industry and Company Machinery Industry - The machinery industry index rose by 0.79% in May, underperforming the CSI 300 index by 1.06 percentage points [8] - The TTM price-to-earnings ratio and price-to-book ratio for the machinery industry are approximately 31.03 and 2.42, respectively, remaining stable month-on-month [8] - The PMI index for May was 49.50%, indicating a 0.5 percentage point increase, reflecting a gradual strengthening of domestic demand [9] - In May 2025, the sales of various excavators reached 18,202 units, a year-on-year increase of 2.12% [9] - Key investment recommendations include companies such as Huace Testing, Guodian Measurement, and XCMG [9][10] Internet Industry - OpenAI launched the o3-pro AI model, enhancing performance and output accuracy [10] - ByteDance released the Doubao large model 1.6, adopting a unified pricing model [10] - The overall performance of the internet sector remains stable, with recommendations for defensive stocks like Tencent Music and NetEase [11] China Petroleum (601857.SH) - China Petroleum has made significant advancements in energy supply and unconventional oil and gas development [12] - The company is focusing on a "reduce oil and increase specialty" strategy through refinery upgrades and extending the natural gas industry chain [13] - Profit forecasts for 2025-2027 are maintained at 167.4 billion, 170.9 billion, and 174 billion, with diluted EPS of 0.91, 0.93, and 0.95, respectively [13]

Investment Rating - The report maintains an "Outperform" rating for the industry, indicating expected performance above the market benchmark by over 10% [3][30]. Core Insights - The overall performance of the internet sector remains stable, with intense competition in the e-commerce industry as platforms continue to offer incentives to merchants and invest in new growth areas like instant delivery [2][26]. - Major players in AI, such as Tencent and Alibaba, are expected to benefit from their business scenarios in cloud computing and advertising, although short-term AI agent developments require further refinement [2][26]. - The report suggests focusing on defensive stocks with stable earnings and low valuations, specifically recommending Tencent Music and NetEase [2][26]. Summary by Sections Industry Investment Rating - The report rates the industry as "Outperform" [3][30]. Product Applications - OpenAI launched the o3-pro AI model, which excels in performance and precision, surpassing competitors in various tests [20]. - Meta introduced the V-JEPA 2 model, enhancing environmental understanding and prediction capabilities [20]. - ByteDance released the Doubao model 1.6, which supports multiple input types and has shown significant improvements in reasoning and mathematical capabilities [21]. Underlying Technologies - Alibaba's Tongyi Lab open-sourced the MaskSearch pre-training framework, enhancing AI's problem-solving capabilities [23]. - DeepMind and Brown University are developing "Force Prompt" technology for realistic motion simulation without 3D models [24]. Industry Policies - The Ministry of Industry and Information Technology discussed strategies to promote AI applications in manufacturing, emphasizing the importance of digital transformation [25]. Key Events Preview - The report highlights upcoming events such as the MWC Shanghai 2025 and the Amazon Cloud Technology Summit [28].

Investment Rating - The report maintains an "Outperform" rating for the industry, indicating expected performance above the market benchmark by over 10% [3][30]. Core Insights - The overall performance of the internet sector remains stable, with intense competition in the e-commerce industry as platforms continue to offer incentives to merchants and invest in new growth areas like instant delivery [2][26]. - Major players in AI, such as Tencent and Alibaba, are expected to benefit from their business scenarios in cloud computing and advertising, although short-term AI agent developments require further refinement [2][26]. - The report suggests focusing on defensive stocks with stable earnings and low valuations, specifically recommending Tencent Music and NetEase [2][26]. Summary by Sections Industry Investment Rating - The report rates the industry as "Outperform" [3][30]. Product Applications - OpenAI launched the o3-pro AI model, which excels in performance and precision, surpassing competitors in various tests [20]. - Meta introduced the V-JEPA 2 model, enhancing environmental understanding and prediction capabilities [20]. - ByteDance released the Doubao model 1.6, which supports multiple input types and has shown significant improvements in reasoning and mathematical capabilities [21][22]. Underlying Technologies - Alibaba's Tongyi Lab open-sourced the MaskSearch pre-training framework, enhancing AI's problem-solving capabilities [23][24]. - DeepMind and Brown University are developing "Force Prompt" technology for realistic motion simulation without 3D models [24]. Industry Policies - The Ministry of Industry and Information Technology discussed strategies to promote AI applications in manufacturing, emphasizing the importance of digital transformation [25]. Key Events Preview - The report highlights upcoming events such as the MWC Shanghai 2025 and the Amazon Cloud Technology Summit [28].

Core Viewpoint - ByteDance's recent launch of the Doubao model 1.6 represents a significant upgrade in its AI model family, enhancing both performance and cost-effectiveness, which is expected to directly benefit AI implementation across various industries [2][4]. Group 1: Product Launch and Features - The Doubao model family has been expanded with the release of Doubao model 1.6, Doubao Video Generation Model Seedance 1.0 Pro, and Doubao Real-time Voice Model, establishing itself as a leading model with full modality, full size, and high cost-performance [2][3]. - Doubao model 1.6 includes three models: 1. Doubao-seed-1.6, an all-in-one model supporting 256K context, deep thinking, multi-modal understanding, and graphical interface operations [3]. 2. Doubao-seed-1.6-thinking, a reinforced version focusing on deep thinking capabilities, with improvements in coding, mathematics, and logical reasoning [3]. 3. Doubao-seed-1.6-flash, a high-speed version with low latency and strong visual understanding capabilities [3]. Group 2: Performance Metrics - By May 2025, the daily token usage of Doubao model exceeded 16.4 trillion, marking a 137-fold increase since its initial release in May 2024 [2]. - Doubao model 1.6 has shown significant improvements in reasoning, mathematics, instruction adherence, and agent capabilities, supporting complex agent construction and enhancing AI productivity [2][3]. Group 3: Pricing Strategy - Doubao model 1.6 introduces a novel pricing model based on input length, with a uniform token price regardless of whether deep thinking mode is activated [4]. - For the most commonly used input range of 0-32K, the pricing is set at 0.8 yuan per million tokens for input and 8 yuan per million tokens for output, representing a 63% cost reduction compared to previous models [4]. - A promotional pricing tier has been established for inputs of 32K and outputs within 200 tokens, further lowering costs to 0.8 yuan per million tokens for input and 2 yuan per million tokens for output [4].

Investment Rating - The report rates the industry as "Outperform" [1][7]. Core Insights - The development of AI foundation models is progressing positively, with decreasing costs, which is expected to accelerate the adoption of AI applications [2][7]. - The report highlights specific companies to watch, including Empyrean Technology, Dameng Database, Beijing Kingsoft Office Software, Newland Digital Technology, Jiangsu Tongxingbao Intelligent Transportation Technology Co., Ltd., Guangzhou Sie Consulting, and Hehe Information, with Wuxi Unicomp Technology Co., Ltd. as a related target [2][7]. Summary by Sections - **AI Model Developments**: OpenAI's o3-pro model was launched, outperforming Google's Gemini 2.5 Pro and Anthropic's Claude 4 Opus in key benchmarks, with an 80% price reduction for the o3 model, enhancing its cost-effectiveness [8]. - **Doubao Model Launch**: The Doubao 1.6 model introduced innovative pricing based on input length, significantly reducing costs by 63% compared to previous models, which is expected to facilitate AI adoption across various industries [9]. - **Ant International's Stablecoin Plans**: Ant International plans to apply for a stablecoin license in Hong Kong, indicating a positive outlook on stablecoin development and potential market growth [10].

Core Insights - The recent FORCE conference in Beijing showcased the launch of new AI models by Volcano Engine, including the Doubao Model 1.6 and Seedance 1.0 pro, highlighting advancements in multimodal interaction and content generation capabilities [2][3] - The global AI model market is highly competitive, with Volcano Engine's new models standing out due to their comprehensive multimodal capabilities and cost-effective pricing strategies [2][3] Product Launches - Doubao Model 1.6 supports multimodal understanding and graphical interface operations, excelling in complex reasoning and multi-turn dialogue tasks, ranking among the top globally [3] - Seedance 1.0 pro generates high-quality 1080P videos with seamless transitions and has achieved top rankings in international assessments for video generation tasks [4] Industry Applications - In the automotive sector, Mercedes-Benz has partnered with Volcano Engine to enhance its smart cabin information retrieval and system response speed using the Doubao model [8] - In finance, Haier Consumer Finance has implemented a tailored large model to meet over 90% of intelligent scenario needs, significantly improving operational efficiency and reducing risks [8] - In education, collaborations with top universities have led to the development of AI applications that enhance research efficiency and quality [9] Technological Innovations - Volcano Engine has introduced an Agent development suite that innovates the entire lifecycle of AI Agent development, enhancing user intent parsing and instruction optimization [5] - The launch of a multimodal data lake solution addresses challenges in data processing, improving resource efficiency and compatibility with various systems [6] - AICC's secure computing technology enhances AI security and privacy, reducing data leakage risks through hardware-protected environments [7] Future Trends - The development of intelligent Agents is expected to drive digital transformation in enterprises, with trends indicating deeper multimodal integration and enhanced autonomous learning capabilities [12][14] - Gartner predicts that by 2028, at least 15% of daily work decisions will be made using Agentic AI, highlighting the growing importance of intelligent Agents in business [12]