作者 | 李笑寅 编辑 | 硬 AI 科技巨头财报季来临，市场似乎很自满，高盛警告仍有风险。 自4月低点以来，标普500指数已上涨26%，主要由科技股推动。未来两周内，科技板块ETFXLK指数中 50%的权重股票将公布业绩。本周三，谷歌母公司Alphabet和特斯拉将率先公布二季度业绩，IBM和德州 仪器等也将陆续登场。 值得警惕的是，高盛数据显示， 当前市场对科技股财报日的预期波动率已降至20年来最低水平，仅为 4.7%，而这种异常的市场平静往往暗示着潜在风险。 分板块看，半导体板块目前是TMT领域最拥挤的投资标的，英伟达、博通、台积电等是最受欢迎的做多标的；软件板块 情绪持续走低，仅微软、甲骨文等头部公司预期向好；互联网板块多空比例处于中位水平，投资者仍关注企业AI应用的 ROI。 硬·AI 微软获得机构持仓集中度评级9分，今年市值增加6500亿美元至接近4万亿美元，成功突破18个月的盘整 格局。 投资者预期其Azure业务本季度增长将达到30%以上。 最受欢迎的软件股做多标的包括微软、Snowflake、甲骨文、ServiceNow和CrowdStrike。做空标的则包 括Adobe、Workday、A ...

Core Viewpoint - Nvidia has informed its Chinese customers that the inventory of the H20 AI chips, customized for the Chinese market, is limited and there are currently no plans to resume production [1][2][3] Group 1: Inventory and Production Status - Nvidia is only fulfilling orders that can be supported by the existing inventory [3] - The production interruption means that even if Nvidia wishes to resume supply, it will face significant time costs, with new chip manufacturing potentially taking nine months from scratch [3] Group 2: Customer Communication and Future Orders - Nvidia is currently communicating with some of its largest Chinese customers to understand their specific chip needs, such as the quantity of H20 or future Blackwell chips they wish to purchase [3] - The company is also collecting customer feedback on the H20 and suggestions for improvements for the next generation of products, which will help determine whether to continue placing additional H20 orders [3]

Core Viewpoint - The article discusses the advancements in semiconductor technology, particularly in AI applications, focusing on key trends such as advanced packaging, CPO technology, and cooling solutions to address performance and efficiency challenges in AI accelerators [2][3]. Advanced Packaging Technology - Advanced packaging is evolving through Chiplet technology and hybrid bonding to enhance AI processor performance. The shift from silicon interposers to silicon bridges and organic RDL is aimed at cost reduction, with a future transition to panel-level packaging expected by 2028-2029 [4][5]. - Hybrid bonding is crucial for improving performance by reducing the bonding area through enhanced alignment precision [5]. CPO Technology - CPO (Co-Packaged Optics) is identified as the next-generation connection technology for AI data center servers, effectively reducing power consumption in high-bandwidth scenarios. However, high costs and the complexity of precise assembly remain significant challenges [6]. - The introduction of next-generation 448Gb SerDes technology may increase CPO adoption, as it addresses signal degradation issues by minimizing transmission distances [6]. Client Device Packaging - In client devices, semiconductor manufacturers are carefully selecting between Chiplet and monolithic architectures based on cost and performance considerations. For instance, AMD's latest Radeon series GPU has integrated previously Chiplet-based SRAM into a monolithic design [7]. - Apple's Vision Pro features a Chiplet package with two high-bandwidth custom DRAM chips, showcasing the trend towards specialized high-performance processors [7]. Cooling Solutions - Traditional cooling methods like air and water cooling are becoming less effective due to increasing power density in AI accelerators. Two-phase liquid cooling is emerging as a key solution due to its high energy efficiency and broad applicability [3][9]. - Different cooling technologies are suited for varying thermal densities: air cooling for below 10W/cm², two-phase liquid cooling for 10-100W/cm², and water cooling for above 100W/cm². The next-generation 3nm AI data center GPUs are expected to have thermal densities around 100W/cm², making two-phase liquid cooling particularly relevant [10][11][12].

Core Viewpoint - Meta is redefining the future of superintelligence with a focus on "personalized super intelligence," aiming to empower billions of users rather than just enhancing enterprise productivity [2][10] Group 1: Investment in AI Infrastructure - Meta is investing thousands of billions in building massive computing clusters, with the largest project, Hyperion, nearing the size of Manhattan [2][3] - The company is constructing multiple gigawatt-scale data centers, with the Prometheus and Hyperion clusters expected to exceed 1 gigawatt, and Hyperion set to expand to 5 gigawatts in the coming years [3][4][15] - Meta's strong business model supports these investments, allowing the company to fund these large-scale projects independently without external financing [4][14][16] Group 2: Talent Acquisition Strategy - Meta is engaged in a fierce competition for top talent, with a focus on hiring 50 to 70 elite researchers to build a high-performing team [4][8] - The company is willing to offer substantial compensation packages, although specific figures reported may not be entirely accurate [4][8][16] - The strategy emphasizes having a small, highly skilled team with maximum GPU resources, which is seen as a strategic advantage in attracting top talent [4][8][16] Group 3: Vision for AI Interaction - Zuckerberg believes AI glasses will become the optimal form of interaction with AI, potentially becoming essential for cognitive enhancement [4][12][13] - These glasses will be capable of observing daily life and providing real-time information, enhancing personal relationships and cultural engagement [12][13] Group 4: Future Outlook on Superintelligence - There are varying opinions on when superintelligence will be realized, with estimates ranging from three to seven years; however, Zuckerberg is optimistic about its potential readiness in two to three years [5][7] - The company aims to leverage its substantial computing power to support the development of superintelligence, which is expected to significantly impact both company operations and broader societal functions [6][10][14]

Core Viewpoint - The strong performance of ASML in Q2 is driven by AI investments, with total revenue reaching €7.7 billion and net profit at €2.3 billion, both at the upper end of guidance. However, management warns of increasing uncertainties due to macroeconomic and geopolitical developments, which may hinder growth in 2026 [1][2][7]. Financial Performance - Q2 net sales amounted to €7.69 billion, exceeding market expectations of €7.51 billion [3]. - Q2 net profit was €2.29 billion, surpassing the market forecast of €2.05 billion [4]. - The order intake for Q2 was €5.54 billion, a 41% increase quarter-over-quarter, with EUV equipment orders at €2.3 billion [5]. - Gross margin reached 53.7%, exceeding expectations, primarily due to high-margin upgrade business and one-time cost reductions [6]. Future Outlook - Despite strong order performance, ASML's management remains cautious about future growth prospects. The CEO indicated that while the fundamentals for AI customers will remain strong in 2026, uncertainties from macroeconomic and geopolitical factors are increasing [7]. - The company expects Q3 net sales to be between €7.4 billion and €7.9 billion, with a gross margin between 50% and 52% [11]. - For the full year 2025, ASML anticipates a revenue growth of approximately 15% and a gross margin of around 52% [12]. Shareholder Returns - ASML announced an interim dividend of €1.60 per share and executed a share buyback of approximately €1.4 billion in Q2 [13].

Core Viewpoint - The resumption of H20 chip sales by NVIDIA to China is expected to have a positive impact on the Chinese internet data center (IDC) industry and related companies, potentially boosting NVIDIA's revenue significantly and benefiting the entire AI semiconductor supply chain [1][4][15]. Group 1: Impact on Chinese IDC Industry - Analysts from Citigroup and Jefferies believe that the restart of H20 chip sales will positively affect the Chinese IDC sector, with a bullish outlook on related stocks [5][4]. - Following the announcement, stocks of major cloud service providers such as Alibaba and Kingsoft Cloud saw significant gains, with Alibaba's Hong Kong shares rising nearly 7% and Kingsoft Cloud's U.S. shares increasing by 18.7% [2][4]. Group 2: Financial Implications for NVIDIA - Bernstein estimates that for every $10 billion in revenue recovered in the Chinese market, NVIDIA's earnings per share (EPS) could increase by approximately $0.25 [9]. - The resumption of sales could help NVIDIA recover a substantial portion of the $15 billion in data center revenue previously at risk, including an anticipated $4-5 billion in revenue for the second half of the year [8][9]. - Melius Research has raised NVIDIA's target price by 43%, projecting that the company's market value could exceed $5 trillion due to the H20 chip sales resumption [1][15]. Group 3: Broader Market Effects - The approval of H20 chip sales is seen as beneficial not only for NVIDIA but also for the entire AI semiconductor supply chain and Chinese tech platforms developing AI capabilities [17]. - The U.S. government's decision to allow NVIDIA to sell H20 chips to China is viewed as a positive development for U.S.-China relations, with implications for ongoing negotiations between the two countries [17][18].

Core Viewpoint - Meta is undergoing a significant strategic transformation to enhance its computational capabilities and compete with leading AI labs like OpenAI, focusing on building large-scale data centers and recruiting top talent [2][12]. Group 1: Infrastructure Development - Meta is launching two massive AI clusters named Prometheus and Hyperion, with Prometheus having a capacity of 1 GW and Hyperion expected to exceed 1.5 GW by the end of 2027, making it the largest single AI data center park globally [1][9]. - The company is adopting a "tent-style" data center design inspired by xAI, prioritizing construction speed and efficiency by using prefabricated power and cooling modules [4][6]. - Meta's strategy aims to transition from being "GPU-poor" to "GPU-rich," enabling it to match the training capabilities of top AI laboratories [6]. Group 2: Strategic Failures and Lessons - The aggressive transformation is partly a response to the failure of Meta's Llama 4 model, which damaged its reputation after the success of Llama 3 [8]. - Key technical failures of Llama 4 included architectural missteps, data quality issues, and challenges in scaling and evaluation, which Meta aims to address through its new initiatives [10][11]. Group 3: Talent Acquisition and Strategic Investments - Meta is focusing on recruiting top talent to bridge the gap with leading AI labs, offering compensation packages that can reach up to $200 million over four years for top researchers [12][13]. - Strategic acquisitions, such as the investment in Scale AI, are seen as crucial steps to enhance Meta's capabilities in data and evaluation, directly addressing the shortcomings revealed by Llama 4 [14][15].

Core Viewpoint - The article emphasizes that true smart hardware should not merely be a connected device but must possess an AI brain capable of deep thinking and understanding, exemplified by the integration of "Wow" companion machine with Quark AI, setting a new industry standard [1][4][30]. Group 1: Definition of Smart Hardware - The article questions the traditional definition of smart hardware, suggesting that simply being connected and mobile does not equate to intelligence [2]. - The "Wow" companion machine demonstrates a transformation in understanding and decision-making, showcasing the potential of AI in enhancing user experience [2][4]. Group 2: Three Essential Elements of Smart Hardware - The first essential element is a smooth interaction interface, which serves as the foundation of intelligence, allowing for natural and hands-free communication [5][9]. - The second element is a powerful cognitive core, enabling the device to engage in multi-modal interactions, providing answers through various media formats [11][12][14]. - The third element is a reliable knowledge system, connecting the device to a robust AI search engine, allowing for real-time and traceable information [17][18][19]. Group 3: Multifaceted Roles of AI Companion - The AI companion serves as an emotional support system, recognizing user emotions and providing comfort through music and meditation [21][23][24]. - It acts as an educational assistant for children, offering creative storytelling and guiding them through learning processes without directly providing answers [25]. - Additionally, it functions as a productivity assistant, helping with planning, decision-making, and information organization across various settings [26]. Group 4: Addressing Modern Life Challenges - The "Wow" companion machine resolves the conflict between fixed spaces and mobile lifestyles, offering a portable entertainment solution that adapts to various living environments [28]. - It bridges the gap between emotional companionship and efficiency, transforming from a passive listener to an active collaborator in users' lives [28][29]. Group 5: Future of Smart Hardware - The integration of Quark AI with the "Wow" hardware signifies a pivotal shift in the evolution of smart devices, where competition will focus on AI capabilities rather than hardware specifications [30][31]. - Devices lacking a deeply integrated AI brain will remain outdated, while those that embrace this evolution will represent the future of intelligent partnerships [31].

Core Viewpoint - Generative AI tools are expected to drastically reduce software development costs, potentially disrupting the high-margin software industry where profit margins exceed 90% [1][15]. Group 1: Impact of Generative AI on Software Development - AI programming tools are significantly lowering the cost and time required for coding, with tasks that previously cost thousands of dollars now achievable for just a few cents [2][12]. - The software industry is experiencing a "peak moment," similar to the disruption seen in traditional media with the rise of the internet and platforms like YouTube [2][4]. - The cost of software development is now only a fraction of what it used to be, with the potential for millions of lines of code to be produced at a fraction of the previous cost [13][14]. Group 2: Comparison with Traditional Media - The decline of traditional media is attributed to the fundamental change in content creation and distribution, with YouTube allowing millions of creators to emerge due to lower entry costs [11][8]. - The number of YouTube channels has skyrocketed, with over 1.139 million channels, compared to a few thousand traditional media channels [11][8]. - The cost to start a YouTube channel is significantly lower than that of launching a traditional TV show, mirroring the trends in software development costs [11][12]. Group 3: Future of the Software Industry - The era of high profits from software is predicted to end, as AI programming agents will lead to an exponential increase in software supply, overwhelming traditional software manufacturers [18][19]. - The traditional software companies' profit margins may erode as the competitive landscape shifts towards sales and marketing costs rather than development costs [18][19]. - The article suggests that while traditional software companies may still have opportunities for profit, this will likely occur through a wave of consolidation in the industry [19]. Group 4: Philosophical Perspective on Software and Hardware - The author posits that software may merely represent a "local minimum" in technological advancement, with the true value lying in hardware [21][24]. - Historically, software was bundled with hardware, and the concept of software as a standalone product emerged only after hardware became widespread [22][23]. - As AI tools enable limitless software generation, the focus may shift back to hardware as the new scarce resource, concentrating value in chips and computational power [24].

Core Insights - Microsoft is experiencing a dual challenge of cost reduction and capital pressure amid its AI transformation, with significant cost savings and revenue growth reported alongside record layoffs and increased infrastructure investment [2][4]. Group 1: AI Cost Savings and Revenue Growth - Microsoft achieved over $500 million in cost savings in its customer service centers through AI tools, while sales personnel using Copilot saw a 9% increase in revenue [2][4]. - AI technology is also impacting engineering, with 35% of the code for new products generated by AI, significantly shortening development cycles [2][4]. - The GitHub Copilot has become a leader in the AI programming tool market, reaching 15 million users by April [2]. Group 2: Layoffs and Capital Expenditure - Microsoft is set to lay off a total of 15,000 employees by 2025, marking the largest layoffs in the company's history, with 9,000 sales positions cut in July [4]. - The backdrop of these layoffs is a surge in AI infrastructure investment, with capital expenditures expected to reach $80 billion over the 12 months ending in June, a 43% increase from the previous year's $56 billion [4]. Group 3: Sales Team Restructuring and AI Adoption - Despite the layoffs, the sales department has shown strong performance, with Azure cloud services and AI Copilot product sales exceeding quarterly targets [5]. - Microsoft is restructuring its sales team to streamline the "solution areas" from six to three, focusing on promoting AI products more effectively [5]. Group 4: Employee Engagement with AI Tools - Microsoft management is actively promoting the use of AI tools among remaining employees, incorporating AI usage into performance evaluations and hosting competitions to encourage productivity improvements [7]. - The emphasis on AI skills is seen as a critical opportunity for employees to invest in their own capabilities, with internal tracking of AI-generated code being implemented [7].