DeepSeek 重构算力基建长期价值的认知 [Table_Industry] 计算机 股票研究 /[Table_Date] 2025.03.14 [Table_Invest] 评级： 增持 上次评级: 增持 | [table_Authors] 李博伦(分析师) | 伍巍(研究助理) | 钟明翰(研究助理) | | | --- | --- | --- | --- | | 0755-23976516 | 021-38031029 | 021-38031383 | | | libolun@gtjas.com | wuwei028683@gtjas.com | zhongminghan029903@gtjas.com | [Table_Report] 相关报告 | | 登记编号 S0880520020004 | S0880123070157 | S0880124070047 | | 本报告导读： 市场低估了 DeepSeek 生态对算力需求的放大效应，我们预计仅其推理端就将产生 近百万 PFLOPS 的算力需求。精度支持及通信效率率先突破的国产 AI 芯片厂商将 获得显著的发展机会。 投资要点： 计算机《AI 应用 ...

Core Insights - The discussion revolves around how value will be created and sustained in the AI-driven era, emphasizing the importance of unique market entry strategies, specialized knowledge, and access to unique data sources [3][4][5] - Companies in sectors like finance, law, and healthcare are highlighted as potential areas for creating lasting value due to their complexity and the foundational work required [3][4] - The balance between showcasing future capabilities and current model limitations is crucial for both startups and established vertical SaaS companies [5][6] Group 1: Value Creation in AI - Unique market entry strategies and specialized knowledge are essential for creating value in the AI landscape [3][4] - Companies that can leverage foundational models while maintaining a deep understanding of their specific industries will thrive [4][5] - Startups may benefit from over-promising during early adoption phases, while established companies face challenges in managing customer expectations [5][6] Group 2: Product Development Challenges - Startups must decide whether to build products based on current technology or anticipated future advancements, as model quality significantly impacts product outcomes [6][7] - The rapid evolution of AI models necessitates a careful approach to product design, balancing speed of release with quality and user experience [19][20] - Companies must develop robust evaluation frameworks to adapt to changing models and user needs, ensuring their products remain relevant [20][21] Group 3: Competitive Landscape - The AI market is becoming increasingly competitive, with numerous companies releasing products simultaneously, complicating product marketing strategies [24][25] - Companies must navigate the complexities of product releases and user expectations, balancing innovation with stability [22][23] - The importance of brand loyalty is emphasized, as users tend to identify with specific models, impacting their long-term engagement [27][28] Group 4: Data and Model Quality - The future of AI models may rely on a combination of human and synthetic data, with the best models emerging from this integration [15][16] - The quality of models is closely tied to the data used for training, highlighting the significance of having strong foundational data sources [30][31] - Companies must focus on the practical application of models in real-world scenarios to demonstrate their value [31][32] Group 5: Global AI Capabilities - There is a recognition that the capabilities of AI in China are often underestimated, with significant advancements being made in the field [32][33] - The emergence of parallel entrepreneurial ecosystems in regions with restricted access to Western platforms has led to innovative solutions [32][33] - Companies must be aware of the global competitive landscape and the potential for new entrants to disrupt established markets [37][38]

Investment Rating - The report maintains an "Outperform" rating for the computer industry [1]. Core Insights - The computer industry is experiencing a slowdown in revenue and profit growth, with software business revenue expected to reach 13.73 trillion yuan in 2024, a 10.0% year-on-year increase, down from 13.4% in 2023 [4][10]. - The report highlights significant capital expenditure increases from major tech companies, indicating a strong investment trend in AI and computing infrastructure [49][52]. Summary by Sections 1. Industry Data - The software industry in China is projected to see a revenue growth of 10.0% in 2024, down from 13.4% in 2023, with total profits expected to grow by 8.7% [4][10][11]. - Software exports are anticipated to increase by 3.5% in 2024, recovering from a decline in the previous year [11]. 2. High-Growth Sectors in 2024 - Integrated Circuit (IC) design is expected to be the highest growth sector, with a projected increase of 16.4% [13]. - Embedded system software is forecasted to grow by 11.8%, driven by ongoing AI advancements [14]. - E-commerce platform services are also expected to grow by 11.4% [15]. 3. Localization - The dependency on imported integrated circuits is at 78%, indicating a 22% localization rate, which has decreased by 2% [20][21]. - Nvidia's revenue from mainland China has decreased, reflecting the impact of U.S. sanctions [23]. 4. AI Developments - The launch of DeepSeek-R1 has intensified competition in the AI model space, with significant advancements in open-source models [25][27]. - DeepSeek's open-source initiative has garnered global attention and is expected to accelerate AI technology development [32][38]. 5. Computing Power - Domestic computing power investments are accelerating, with major tech firms planning substantial capital expenditures [49][52]. - Nvidia's new Blackwell chip has significantly contributed to its revenue growth, indicating strong demand for advanced computing solutions [55][56].

Core Viewpoint - The article analyzes the historical development of the internet since the 1990s and the Dot-com bubble, drawing parallels to the current trends in AI development, suggesting that understanding past trends can provide insights into future industry and market dynamics [1][7]. Industry Perspective - The challenge lies in grasping the "timing" and "development path" of the industry. While the trends in the internet industry can be anticipated, accurately pinpointing the timing and specific forms of development is challenging. For instance, the World Wide Web and PCs were not initially mainstream forms [3][19]. - The early internet's core features included open cooperation, network effects, and decentralization, which ultimately shaped its evolution. The transition from localized networks to a unified internet infrastructure was not initially predictable [11][12]. - The early internet's leading companies leveraged their resource advantages to dominate the market, a trend that may re-emerge in the current AI landscape [19]. Market Perspective - The Dot-com bubble was a culmination of a long bull market in the U.S., with significant growth in internet penetration from 0% to 30% between 1990 and 1998. This period saw a surge in IPOs for internet-related companies [20][34]. - The valuation logic for companies shifted during the bubble, with non-rational factors dominating market trends. After the bubble burst, the market returned to fundamentals, leading to a significant drop in bandwidth costs by 90% and a talent surplus in computing [20][29]. Insights - The current AI trend is seen as entering an application phase, with the ultimate goal being AGI (Artificial General Intelligence). However, there is no consensus on the path or timeline to achieve this [4][36]. - The emergence of open-source AI technologies like DeepSeek is likened to the early internet's transition to open applications, potentially democratizing access to AI capabilities [38][45]. - The article suggests that the current AI development phase may mirror the early internet era, where initial applications are being developed, and the market is still defining its standards and models [39][41]. Conclusion - The historical analysis indicates that while identifying major trends is relatively straightforward, determining the timing and specific forms of development is complex. The interplay of necessity and randomness plays a crucial role in shaping industry trajectories [19][34]. - The article emphasizes that the aftermath of the Dot-com bubble laid the groundwork for sustainable business models and infrastructure, which could similarly apply to the current AI landscape as it matures [35][42].

Group 1 - Google has introduced new personalization features for its Gemini chatbot, allowing it to reference users' Google Search histories for better recommendations, which is an opt-in feature [1] - Users can now connect various apps such as Calendar, Notes, Tasks, and Photos to Gemini, enhancing its functionality [2] - The company aims to strengthen its position in the competitive AI industry, with a focus on scaling Gemini for consumer use in the upcoming year [3] Group 2 - Google launched open-source Gemma 3 models for developers, capable of analyzing text, images, and short videos, which the company claims to be "the world's best single-accelerator model" [4] - New AI models, Gemini Robotics and Gemini Robotics-ER, were introduced, both operating on Gemini 2.0, which is described as the company's "most capable" AI to date [6]

Core Insights - Alibaba Group launched a new version of its AI assistant app powered by its Qwen AI reasoning model to enhance its competitive edge in the AI application market [1][3] - The updated app integrates various functions such as chatbot capabilities, deep thinking, and task execution into a single platform [2] - Alibaba's chairman emphasized the importance of practical applications in maximizing AI model intelligence [2] Investment and Development - Alibaba unveiled its latest AI reasoning model, QwQ-32B, claiming it rivals leading models like DeepSeek-R1 [3] - The company plans to invest 380 billion yuan ($52.5 billion) in cloud computing and AI infrastructure over the next three years [3] - Qwen AI has reportedly performed well in official benchmark tests, indicating Alibaba's growing influence in the AI sector [3] Partnerships and Market Position - Manus AI, developed by the startup Butterfly Effect, announced a strategic partnership with Alibaba, aiming to outperform OpenAI's DeepResearch [4] - Experts noted that Alibaba is making significant progress in its AI cloud business, with a notable profit increase in the December quarter driven by its Cloud Intelligence unit and e-commerce segment [5] - Alibaba has secured a partnership with Apple Inc for AI integration on iPhones, positioning itself to compete with OpenAI [5] Market Performance - Alibaba's shares in Hong Kong fell by 2.45% to 131.5 Hong Kong dollars ($16.9) on the day of the news [5]

Group 1 - The overall forward P/E ratio and median of the US stock market reached the highest level in 22 years (excluding the pandemic bubble period) by early 2025, indicating extreme market optimism and a significant valuation bubble, particularly among the "Seven Sisters" (Apple, Microsoft, Amazon, Nvidia, etc.), whose market capitalization accounts for nearly 60% of US GDP, far exceeding reasonable levels [1][2] - The proportion of financial assets allocated to stocks by US households reached a historical high of 43.4%, indicating that market risk tolerance has reached its limit, with excessive capital concentration in tech giants leading to high liquidity dependence on a few companies [2] Group 2 - DeepSeek's innovative MLA architecture and MoE Sparse structure reduced model training costs to 5% of that of international giants, with inference capabilities comparable to top models like GPT-4o, undermining Nvidia's chip scarcity and directly impacting the core profit logic of the US AI industry, which relies on high capital investment to create barriers [3][4] - The global competitive landscape is being restructured as DeepSeek demonstrates that computational power embargoes do not constitute absolute barriers, weakening the US's technological advantage and shifting global AI discourse from a US-centric model to a more diversified competition [4] Group 3 - Trump's signing of a memorandum on tariffs related to digital taxes and the imposition of tariffs on Mexico and Canada raised concerns about global supply chain stability, causing companies like Apple to commit to shifting their supply chains to avoid tariff impacts, resulting in a decline of over 6.6% in the S&P 500 and over 10% in the Nasdaq [5][6] - The contradiction in Trump's "America First" policy, which aims to rebuild manufacturing while relying on robots to replace labor, has led to rising unemployment and structural imbalances in job creation, further complicating market expectations and increasing risk aversion [6][7] Group 4 - The failure of Trump's $500 billion AI investment plan (Stargate) due to DeepSeek's low-cost path has diminished global capital confidence in the US AI industry, compounded by rising credit risk indicators in the US, leading investors to worry that policy uncertainty will undermine economic fundamentals [7][8] - The recent Nasdaq decline is attributed to a confluence of factors: the fragile valuation bubble of tech giants under interest rate risks, the direct disruption of AI valuation logic by DeepSeek's technological breakthrough, and the exacerbation of market uncertainties by Trump's policies, prompting a shift of funds from overvalued tech stocks to defensive sectors [8]

Investment Rating - The report does not explicitly state an investment rating for the industry Core Insights - The report discusses the evolution of intelligence and the new normal of human-machine collaboration, emphasizing the transformative impact of AI on various sectors [1][55] - It highlights the significant advancements in AI models, particularly the transition from GPT-3 to DeepSeek-V3, showcasing improvements in training data volume and model architecture [4][6] - The report notes the rapid growth of AI tools and applications, indicating a shift towards more integrated and efficient AI solutions across industries [71][74] Summary by Sections 1. Evolution of Intelligence - The evolution of AI is marked by increasing data volumes and model complexities, with DeepSeek-V3 utilizing 14.8 trillion tokens compared to GPT-3's 300 billion tokens [6] - The report outlines the historical context of AI development, linking it to broader industrial revolutions and technological advancements [64][66] 2. Human-Machine Collaboration - The report emphasizes the importance of human-machine collaboration, suggesting that AI will augment human capabilities rather than replace them [55][57] - It discusses the potential for new job creation alongside job displacement, highlighting the need for skill enhancement in the workforce [57][58] 3. Industry Status - The report provides an overview of the current state of AI applications in various sectors, including consumer and enterprise-level integrations [74] - It notes the deployment of advanced AI models in critical areas such as energy, healthcare, and governance, showcasing their practical benefits [74] 4. Educational Growth - The report stresses the need for educational initiatives to prepare the workforce for the AI-driven future, focusing on skill development and adaptability [57][58] - It suggests that AI can lead to improved work-life balance, potentially enabling shorter workweeks as productivity increases [57][58]

Core Viewpoint - The article discusses the significant drop in Nvidia's market value due to the emergence of China's DeepSeek, which raises concerns about a potential valuation bubble in the AI industry, drawing parallels to the internet bubble of the early 2000s [4][5][6]. Group 1: Nvidia's Market Performance - On January 27, 2025, Nvidia experienced a record single-day drop of 17%, losing approximately $600 billion in market value, marking the highest single-day loss for a company in U.S. stock market history [4]. - Nvidia's stock had previously surged by 239% in 2023 and an additional 171% in 2024, leading to investor skepticism regarding whether its valuation had peaked [4]. Group 2: DeepSeek's Impact - DeepSeek's launch of a low-cost, open-source model, DeepSeek-R1, is seen as a catalyst for Nvidia's market decline and highlights the competitive pressure from Chinese companies in the global AI landscape [4][6]. - The technological advancements and cost control demonstrated by DeepSeek are intensifying global competition and prompting a reevaluation of AI industry valuations [3][4]. Group 3: Market Reflections and Comparisons - The article draws a comparison between the current AI investment frenzy and the internet bubble of 25 years ago, suggesting that both phenomena exhibit similar market dynamics and risks [5][6]. - Historical context is provided, noting that the NASDAQ index reached a peak of 5048 points in March 2000, followed by a significant market correction that resulted in a loss of $6.5 trillion in market value over two years [4][5].

Core Viewpoint - The emergence of DeepSeek from Hangzhou signifies a shift in the innovation landscape, challenging the dominance of major players like OpenAI and recalibrating the innovation strength rankings among Chinese cities [2][8]. Group 1: Comparison of Innovation Between Shenzhen and Hangzhou - Shenzhen has a higher R&D intensity at 6.46% of GDP compared to Hangzhou's 3.92% for 2024 [6]. - Shenzhen has accumulated 1.844 million invention patents over the past decade, significantly more than Hangzhou's 850,000 [6]. - Shenzhen boasts over 25,000 national high-tech enterprises, with a density of 12 per square kilometer, far exceeding Hangzhou's 16,500 [6]. - Despite Shenzhen's advantages in R&D and technology enterprise density, Hangzhou has successfully birthed the "Six Little Dragons," indicating a strong innovation atmosphere [6][8]. Group 2: Structural Differences in Industry - Shenzhen's industrial structure is heavily focused on manufacturing, with a 2023 GDP contribution of 0.1% from the primary sector, 37.6% from the secondary sector, and 62.3% from the tertiary sector [10]. - In contrast, Hangzhou's GDP contribution is 1.7% from the primary sector, 28.3% from the secondary sector, and 70% from the tertiary sector, highlighting a more service-oriented economy [10]. - Shenzhen's industrial output is primarily in electronics, with significant production in robots, smartphones, and integrated circuits [11][12]. Group 3: Innovation Ecosystem and Strategy - Hangzhou's strategic pivot towards modern service industries has fostered a digital economy, with a projected core industry value of 630.5 billion yuan in 2024, growing at 7.1% [16]. - The city has effectively utilized its private sector's dynamism to develop information services, cultural creativity, and financial services, creating a unique innovation ecosystem [16]. - The contrasting focus of Shenzhen on hardware innovation versus Hangzhou's emphasis on platform economy and algorithm innovation illustrates their distinct paths in the innovation landscape [12][14]. Group 4: Broader Implications of Innovation - The rise of DeepSeek challenges the notion that only hardware breakthroughs constitute true innovation, emphasizing the importance of algorithm optimization and digital economy advancements [18][19]. - The innovation spectrum encompasses both tangible and intangible advancements, suggesting that both Shenzhen and Hangzhou contribute uniquely to China's technological landscape [19][20]. - The competition between Shenzhen and Hangzhou should be viewed as complementary rather than adversarial, as each city plays a distinct role in the broader innovation ecosystem [20].