每经记者｜朱成祥 每经编辑｜陈俊杰 站在AI（人工智能）发展的长河中，2025年可能是非常重要的节点。 2025年，被认为是AI走向大规模应用的开始，是AI智能体的元年。随着AI应用爆发，算力芯片的需求 逻辑也被重塑。推理而不是训练，将成为未来算力需求的核心增长点。 此外，人形机器人的发展也将助推对算力芯片的需求。人形机器人分为大脑、小脑和本体，而算力芯片 正是人形机器人大脑的计算核心。 在WAIC 2025上，各大厂商带来了它们的解决方案。比如华为昇腾的384超节点，摩尔线程"AI工厂"理 念，施耐德电气"算电协同"三层架构等。 华鲲振宇副总裁宋璇表示："AI产业中，我们定位为'国产算力生态的技术转化者'与场景落地者，华鲲 振宇不仅要发展积累AI产品能力，更要坚定地投入到国产AI生态建设中，我们深耕鲲鹏+昇腾生态，通 过与华为在服务器领域深度协同，将生态技术红利精准输送到千行百业。目前我们已实现整机出货量第 一，在金融、运营商、政府等领域积累了深厚实践经验。" 除了华为昇腾这类NPU，在当下火热的GPGPU（通用图形处理器）赛道，国产厂商也带来了各自的产 品。其中摩尔线程以全功能GPU为核心的"云边端"全栈 ...

Core Viewpoint - The release of Kimi K2 marks a significant turning point for the company, indicating a shift from a reliance on scaling laws to a more innovative approach in AI model development and strategy [2][4][22]. Group 1: Kimi K2 Release and Its Impact - Kimi K2 achieved a global fifth ranking in the LMArena leaderboard and first among open-source models, surpassing competitors like Claude 4 and DeepSeek-R1-0528 [2]. - The release is seen as more than just a temporary success; it represents a deeper strategic shift for the company and the industry [4][22]. - Kimi K2 introduces two major advancements: an expansion of model parameters to over 1 trillion and the concept of "model as agent," allowing for tool utilization [23][35]. Group 2: Challenges Faced by Kimi - Kimi's previous strategy relied heavily on scaling laws, believing that larger models and more data would lead to better performance, but this approach faced challenges as high-quality data became scarce [8][13][14]. - The company's user growth strategy was questioned after competitors like DeepSeek demonstrated significant user acquisition without marketing spend, highlighting the need for a more effective product [18][54]. - Kimi's marketing budget reached approximately 900 million RMB in 2024, yet user engagement declined, indicating a disconnect between spending and user retention [17]. Group 3: Strategic Transformation - The company has shifted its focus from aggressive marketing to enhancing model performance and embracing open-source collaboration, reflecting a significant cultural change [55]. - Kimi's team has decided to halt all marketing activities and concentrate resources on foundational algorithms and the K2 model, emphasizing the importance of product quality over quantity [55]. - The strategic pivot is seen as a response to the success of DeepSeek, which has prompted Kimi to adopt more effective architectural choices and prioritize technical research [55][56].

Investment Rating - The report maintains an "Outperform" rating for the AI application industry [1] Core Insights - The capabilities of AI models are rapidly improving, driven by open-source initiatives that lower costs. Large models have achieved new heights in knowledge Q&A, mathematics, and programming, surpassing human-level performance in various tasks. The introduction of high-performance open-source models like Llama 3.1 and DeepSeek R1 has narrowed the gap between open-source and closed-source models [2][5] - AI agents are becoming more sophisticated, with a surge in new product releases. These agents can perceive their environment, make decisions, and execute actions, enhancing their functionality through the integration of external tools and services [2][30] - The commercial use of AI is on the rise, with significant growth in usage and performance of domestic models. The gap between top models in China and the US is closing, supported by a continuous increase in global AI model traffic [2][50] - AI applications are reshaping traffic entry points, with traditional internet giants leveraging proprietary data and user engagement to integrate AI functionalities into existing applications [2][50] - The open-source movement is increasing investment willingness and accelerating cloud adoption among enterprises, as the proliferation of development tools lowers industry application barriers [2][50] Summary by Sections Model Layer: Rapid Capability Enhancement and Cost Reduction - The mainstream model architecture is shifting towards MoE, allowing for more efficient resource use while enhancing performance. Models like DeepSeek-V3 and Llama 4 have demonstrated low-cost, high-performance capabilities [8][9] - The multi-modal capabilities of models have significantly improved, enabling them to process various data types, thus expanding application scenarios [8][9] - The introduction of chain-of-thought reasoning techniques has improved the accuracy and reliability of model responses [8][9] Commercialization: Continuous Growth in Usage and Strong Performance of Domestic Models - The competition among vendors has led to a significant decrease in inference costs, benefiting application developers and end-users [21][22] - The API call prices for major models have dropped substantially, with some models seeing reductions of up to 88% [21][22] AI Agents: Technological Advancements and Product Releases - AI agents are evolving from traditional models to more autonomous entities capable of independent decision-making and task execution [30][31] - The introduction of protocols like MCP and A2A is enhancing the capabilities and interoperability of AI agents, facilitating complex task execution across different systems [38][39] C-end Applications: AI Empowering Business and Reshaping Traffic Entry - AI applications are expected to redefine traffic entry points, with major players actively positioning themselves in this space [2][50] B-end Applications: Open-source Enhancing Investment Willingness and Cloud Adoption - The development of open-source tools is significantly lowering the barriers for industry applications, accelerating the intelligent transformation of various sectors [2][50]

Investment Rating - The report maintains an "Overweight" rating for the computer industry, expecting the industry index to outperform the market index by more than 10% in the next six months [31]. Core Insights - The computer industry index has increased by 4.98% over the past two weeks, outperforming the CSI 300 index by 3.31 percentage points, ranking 4th among 31 first-level industries [10][2]. - The SW computer sector's PE TTM (excluding negative values) is 53.97 times, positioned at the 87.27% percentile over the past five years and 74.59% over the past ten years [20][2]. - The release of Grok 4 by xAI is expected to enhance AI application development, with significant implications for AI computing power and investment opportunities [27][21]. Summary by Sections 1. Industry Performance Review - The SW computer sector has shown a cumulative increase of 11.68% this year, outperforming the CSI 300 index by 9.15 percentage points [10][2]. - The top-performing stocks in the computer sector over the past two weeks include Information Development, Puling Software, and Borui Data, with increases of 46.00%, 42.52%, and 41.85% respectively [16][2]. 2. Valuation Situation - As of July 17, 2025, the SW computer sector's PE TTM stands at 53.97 times, indicating a high valuation relative to historical performance [20][2]. 3. Industry News - Key developments include the launch of Grok 4, which is positioned as a leading AI model, and NVIDIA's resumption of H20 chip supplies to China [21][27]. - Google plans to invest $25 billion in AI infrastructure over the next two years, highlighting the growing demand for AI capabilities [21][27]. 4. Company Announcements - Notable announcements include Star Ring Technology's plan to issue H shares and list on the Hong Kong Stock Exchange, aiming to enhance competitiveness and brand image [24][2]. 5. Weekly Perspective - The report emphasizes the potential of Grok 4 to drive advancements in AI applications, suggesting a focus on investment opportunities in AI computing power and related sectors [27][2]. 6. Recommended Stocks - Suggested stocks for attention include: - GuoDianYunTong (002152.SZ) for its stable growth in fintech and deepening layout in data elements and computing power [29]. - Shenzhou Digital (000034.SZ) as a core partner in the "Kunpeng + Ascend" industrial chain, expected to benefit from rising domestic computing power demand [29].

Core Insights - Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has raised $2 billion in seed funding led by a16z, achieving a valuation of $12 billion, marking it as the largest seed funding round in tech history [1][2] - The initial funding target was $1 billion with a valuation of $9 billion, but the final amount increased significantly over a few months [1] - The company currently lacks specific product offerings and revenue, with only a high-profile founding team and vague technological direction publicly available [1] Company Overview - Mira Murati has been with OpenAI since 2016, serving as CTO and leading the development of groundbreaking technologies like GPT-3, GPT-4, DALL-E, and ChatGPT [2] - The founding team includes notable AI experts such as John Schulman, Barret Zoph, Bob McGrew, Alec Radford, Alexander Kirillov, Jonathan Lachman, and Lilian Weng, all of whom have significant contributions to AI advancements [4][5][7][9][12][13][15] Talent Acquisition in AI Industry - The competition for top AI talent has intensified, with companies like Anthropic, Safe Superintelligence, and Thinking Machines Lab emerging as key players, all led by elite AI researchers [17] - The trend indicates that talent is becoming the most critical factor in the AI industry, surpassing computational power and data [17] - Major tech companies are aggressively acquiring talent, as seen in Meta's recruitment efforts, which include significant investments and hiring from various AI firms [18][19][20] Future Product Development - Thinking Machines Lab plans to release its first product within months, focusing on open-source components and AI solutions tailored to business KPIs, referred to as "reinforcement learning for businesses" [16] - The company emphasizes multimodal capabilities and effective safety measures for AI systems, aligning with industry trends towards responsible AI development [16]

Investment Rating - The report maintains an "Outperform" rating for the communication industry, indicating an expected performance exceeding the benchmark index by over 10% [1][36]. Core Insights - The communication industry has seen significant advancements with the release of Grok4 and Kimi K2, which are expected to enhance capabilities in various applications such as programming and robotics [3][15]. - The earnings forecasts for major players in the server, optical module, and copper connection sectors are promising, with notable year-on-year growth expected [5][16]. - The ongoing global competition in computing power is shifting from model training to service quality and competitive advantages, suggesting a robust outlook for investments in the sector [7][18]. Summary by Sections Industry Investment Rating - The communication industry is rated as "Outperform," with expectations of exceeding the benchmark index by over 10% [1][36]. Industry Trends - Grok4, launched by xAI, boasts a tenfold improvement in reasoning capabilities compared to its predecessor, with applications in complex task execution and programming [3][14]. - Kimi K2, a new MoE model, has achieved state-of-the-art results in several foundational tests, indicating significant advancements in AI capabilities [4][15]. Earnings Forecasts - Industrial Fulian anticipates a net profit of 11.96-12.16 billion yuan for the first half of 2025, reflecting a year-on-year increase of 36.8%-39.1% [5][16]. - Other companies like Guangxun Technology and Huagong Technology also project substantial profit growth, with increases ranging from 30% to 95% year-on-year [5][16]. Investment Recommendations - The report suggests focusing on both overseas and domestic computing power chains, highlighting companies such as Industrial Fulian and Huagong Technology as key players [8][19]. - The ongoing arms race in computing power is expected to yield numerous investment opportunities in the coming years, particularly in the context of domestic algorithm optimization [17][18]. Market Overview - The overall market showed positive performance during the week of July 7-11, 2025, with notable increases in various indices, including a 2.36% rise in the ChiNext Index [8][19]. - Specific sectors such as equipment manufacturers and IoT led the weekly gains, indicating strong investor interest [8][19].

就在几天前，马斯克的xAI正式发布Grok 4大模型，号称世界最强AI。 我们团队这几天仔细研究了Grok 4相关的研究资料，有一些新发现，对未来AI产业趋势及算力展望具有一定价值，遂整理成此 文，用一篇文章的篇幅给大家介绍清楚Grok 4的发展脉络。 核心要点： 下面我们正式开始。 一、大力出奇迹，性能登顶各大Benchmark Grok 4是在xAI自研的Colossus超算上训练而成的，其训练规模远超前代模型，计算资源投入为 Grok-2 的100倍、Grok-3 的 10 倍， 实现了推理性能、多模态能力和上下文处理能力的跃升。 Grok 4拥有两个版本：Grok 4（月费30美金）、Grok 4 Heavy（月费300美金，是的你没看错，300美金！）。其中Grok 4是单Agent 版本，而Heavy是多Agent协作版本，能够同时启动多个Agent并行工作，并最后整合结果。 经过实测，Grok 4在多个Benchmark上均取得了优秀的成绩。在GPQA、AIME25、LCB（Jan-May）、HMMT25、USAMO25等多 项测评中，Grok 4都超越了o3、Gemini 2.5 Pro、Cl ...

Core Viewpoint - The breakthrough in embodied intelligence will not occur in laboratories but in practical applications, indicating a shift from academic research to entrepreneurial ventures in the field [1][5]. Company Overview - Qianxun Intelligent was founded by Gao Yang, a chief scientist and assistant professor at Tsinghua University, and Han Fengtao, a veteran in the domestic robotics industry, to explore the potential of embodied intelligence [2][3]. - The company recently demonstrated its new Moz1 robot, capable of performing intricate tasks such as organizing office supplies [4][3]. Industry Trends - The development of embodied intelligence is currently at a critical scaling moment, similar to the advancements seen with large models like GPT-4, but it may take an additional four to five years for significant breakthroughs [2][29]. - There is a notable difference in the development of embodied intelligence between China and the U.S., with China having advantages in hardware manufacturing and faster repair times for robots [6][7]. Research and Development - Gao Yang transitioned from autonomous driving to robotics, believing that robotics offers more versatility and challenges compared to specialized applications like self-driving cars [10][12]. - The field of embodied intelligence is experiencing a convergence of ideas, with many previously explored paths being deemed unfeasible, leading to a more focused research agenda [12][13]. Technological Framework - Gao Yang defines the stages of embodied intelligence, with the industry currently approaching Level 2, where robots can perform a limited range of tasks in office settings [17][18]. - The preferred approach in the industry is end-to-end systems, particularly the vision-language-action (VLA) model, which integrates visual, linguistic, and action components into a unified framework [19][20]. Data and Training - The training of VLA models involves extensive data collection from the internet, followed by fine-tuning with real-world operation data and reinforcement learning to enhance performance [23][24]. - The scaling law observed in the field indicates that increasing data volume significantly improves model performance, with a ratio of 10-fold data increase leading to substantial performance gains [27][28]. Market Dynamics - The demand for humanoid robots stems from the need to operate in environments designed for humans, although non-humanoid designs may also be effective depending on the application [33][34]. - The industry is moving towards a model where both the "brain" (AI) and the "body" (robotic hardware) are developed in tandem, similar to the automotive industry, allowing for specialization in various components [39][41].

Core Viewpoint - The article discusses the evolution of artificial intelligence (AI) and its intersection with brain science, emphasizing the importance of large models and the historical context of AI development, particularly during its "winters" and the lessons learned from past mistakes [5][18][27]. Group 1: Historical Context of AI - AI experienced significant downturns, known as "AI winters," particularly from the late 1990s to the early 2000s, which led to a lack of interest and investment in the field [2][3]. - Key figures in AI, such as Marvin Minsky, expressed skepticism about the future of AI during these downturns, influencing others like Liu Jia to pivot towards brain science instead [3][14]. - The resurgence of AI began around 2016 with breakthroughs like AlphaGo, prompting a renewed interest in the intersection of brain science and AI [3][14]. Group 2: Lessons from AI Development - Liu Jia reflects on his two-decade absence from AI, realizing that significant advancements in neural networks occurred during this time, which he missed [14][15]. - The article highlights the importance of understanding the "first principles" of AI, particularly the necessity of large models for achieving intelligence [22][27]. - Liu Jia emphasizes that the evolution of AI should not only focus on increasing model size but also on enhancing the complexity of neural networks, drawing parallels with biological evolution [24][25]. Group 3: Current Trends and Future Directions - The article discusses the current landscape of AI, where large models dominate, and the importance of scaling laws in AI development [27][30]. - It notes the competitive nature of the AI industry, where advancements can lead to rapid obsolescence of existing models and companies [36][39]. - The article suggests that future AI development should integrate insights from brain science to create more sophisticated neural networks, moving beyond traditional models [25][50].

Core Insights - The article discusses the advancements in AI, particularly focusing on the evolution of the Transformer model and the introduction of the 2-simplicial Transformer, which enhances the efficiency of token utilization and model scalability [1][4][10]. Group 1: Transformer and AI Development - The paper "Attention Is All You Need" marked a significant turning point in AI development, establishing the Transformer as the foundational paradigm for current language models [1]. - The citation count for this paper is approaching 190,000, indicating its profound impact on the field [2]. - The ongoing challenge in AI is acquiring a sufficient quantity of high-quality tokens and efficiently utilizing them, necessitating further upgrades to the Transformer model [3]. Group 2: 2-Simplicial Transformer - Meta's recent research introduced a rotationally invariant trilinear attention mechanism, demonstrating comparable representational capacity to the 2-simplicial Transformer and potentially altering the coefficients in the Scaling Law [4][10]. - The 2-simplicial Transformer, derived from Clift et al. (2019), generalizes the dot-product attention mechanism to a trilinear form, enhancing its scalability under token constraints [19][11]. - Experimental results indicate that the 2-simplicial Transformer can more effectively approximate the irreducible entropy of natural language compared to traditional dot-product attention Transformers [11]. Group 3: Scaling Law and Model Performance - The Scaling Law describes how loss decreases with the total number of model parameters and token count, suggesting that larger models should approach the irreducible loss of natural text distribution as both parameters and tokens increase [13][15]. - Hoffmann et al. (2022) found that the optimal number of parameters and dataset size should scale proportionally with the computational budget, with estimated scaling exponents around 0.49 for parameters and 0.5 for tokens [17][18]. - The 2-simplicial Transformer exhibits a steeper scaling slope compared to the dot-product attention Transformer, indicating a higher exponent in its Scaling Law [50]. Group 4: Experimental Results - The team conducted experiments with various models, revealing that the 2-simplicial attention mechanism did not provide benefits in models with fewer than 2 billion active parameters [45]. - The performance metrics across different model sizes showed slight improvements or declines when comparing the 2-simplicial Transformer to traditional Transformers, with variations in performance percentages noted [43][44]. - The study estimated the differences in scaling coefficients between the 2-simplicial and dot-product attention mechanisms, highlighting the potential for improved efficiency in larger models [46][49].