扫码成为巴伦创始会员，解锁万字全文！ 美股科技板块正由AI强势引领，2025年第二季度，微软、Meta等四大巨头营收双位数增长，七大科技 龙头盈利预期上修近15%，英伟达、微软等产业链不同环节企业盈利亮眼，AI驱动的增长韧性备受关 注。同时，企业端付费意愿和真实转化能力尚未验证，基础设施瓶颈、高昂的资本开支、监管变数以及 预期与技术的巨大鸿沟，都可能成为引发市场回调的导火索。未来高增长能否延续？AI产业链盈利红 利如何扩散？在第3期《巴伦菁英月谈会》上，华映资本海外合伙人邱谆、上海纽约大学实践助理教授 Nicole Wang，与巴伦中文网高级编辑孙骋，一同直播讨论。 ...

机器之心报道 机器之心编辑部 在 AI 领域，大家通常采取后训练方式来让模型获取专项技能。然而后训练一般依赖带有标注参考的监督微调，或通过可验证的程序化检查器提供奖励。 这就带来一些问题，目前许多有价值的任务可能同时缺乏这两种资源。例如在不可验证的场景中（临床、自由对话和创意写作），可能存在多个有效答案，确定 性规则检查难以实施。 在这种情况下，实践者往往只能依赖（i）繁琐的标注流程，或（ii）通过另一个 LLM 对自由形式输出进行粗略奖励。 然而，当后训练缺乏真实标注时，学习信号从何而来？ 为了回答这一问题，来自牛津大学、Meta 超级智能实验室等机构的研究者提出设想： 推理计算是否可以替代缺失的监督？ 本文认为答案是肯定的，他们提出了一种名为 CaT（Compute as Teacher） 的方法，核心思想是把推理时的额外计算当作教师信号，在缺乏人工标注或可验证答 案时，也能为大模型提供监督信号。 结果显示，推理时直接应用 CaT显著提升了 Gemma 3 4B、Qwen 3 4B 和 Llama 3.1 8B 的性能，即使在不可验证领域（MATH-500 最高提升 27%；HealthBench 提升 ...

Core Insights - The lack of high-quality data has become a bottleneck for the continuous learning and capability enhancement of large language models (LLMs) [1] - Meta has proposed a new reinforcement learning method called "Language Self-Play" (LSP) to enable models to self-improve without relying on additional data [1][2] Methodology - LSP utilizes a self-play framework, treating the model's capabilities as performance in competitive games, allowing it to generate stronger strategies through self-competition [2] - In the LSP framework, the same pre-trained LLM assumes two different roles: "Challenger" generates challenging queries, while "Solver" responds to these queries to maximize task rewards [3][5] Technical Innovations - LSP incorporates two core technologies: - Group Relative Policy Optimization (GRPO) allows the Challenger to generate multiple queries, and the Solver to provide various responses, establishing a quality evaluation benchmark [5] - KL Divergence Regularization prevents the model from deviating too far from the initial reference model, ensuring training effectiveness [5] Evolution of LSP - The initial version, LSP-Zero, relied solely on adversarial interactions but faced issues like "meaningless adversarial games" [6][7] - The upgraded LSP introduces a self-reward mechanism, where a reference model scores the quality of "Challenger query + Solver response," promoting high-quality interactions [7] Experimental Validation - Experiments using AlpacaEval benchmark demonstrated that LSP and LSP-Zero significantly improved the performance of the base model Llama-3.2-3B-Instruct, achieving comparable results to GRPO despite lacking training data [10][11] - LSP outperformed LSP-Zero, particularly in tasks requiring conversational prompts, showcasing its advantages in specific scenarios [11][14] Limitations - LSP showed slightly lower performance than GRPO in the Koala dataset, attributed to the structured nature of queries generated by LSP, which did not align well with the dataset's loose conversational style [16] Future Implications - The introduction of LSP addresses the data dependency issue in large model training and validates the feasibility of "data-free training," potentially reducing training costs and resource investments [17] - The self-play framework may exhibit significant potential for knowledge expansion once AI can collect its own experiential data [17]

环球网消息，美国总统特朗普19日签署公告，将企业为H-1B签证申请人支付的费用提高至10万美元，并称 此举旨在确保美国引进的是高技能、不能被美国人替代的人才。此前H-1B签证费用通常为数千美元。这笔 巨额资金将对亚马逊、微软、Meta、谷歌等科技巨头造成直接冲击，它们一直依赖H-1B项目雇用包括软件 开发工程师在内的外国员工。在刚过去的周末，美国企业陷入一片混乱，紧急叫回正在海外旅行的受影响 员工，评估政策调整产生的影响。尽管白宫20日对新规的适用人群等具体内容作出进一步解释，但有人形 容，整个政策依旧"处处都是问号"。提高H-1B签证费又将对美国的企业竞争力、创新能力、经济增长造成 怎样的长期影响？美国舆论对于这一问题的答案充满不安。与此同时，这项新规又给近期磕磕绊绊的美印 关系增添新问题：按国籍看，印度人是去年获得H-1B 签证最多的人群，占比超过70%。 据印媒21日报道，在一架从旧金山起飞的国际航班上，关于H-1B签证的消息引发机舱内印度旅客的慌乱。（《印 度时报》网站） 如同"在雾中飞行" 根据相关规定，H-1B签证允许美国雇主在找不到合格美国雇员的情况下，雇用具有专业技能的外国员工。 持有者最长可 ...

Core Insights - The Meta Connect conference is a significant annual event for global tech professionals, where new technologies and products are showcased [1] - Meta's new AI glasses are highly anticipated, especially in light of the current challenges faced by the AI glasses market, which is struggling with quality and consumer satisfaction [2][4] - The release of three new AI glasses models by Meta aims to address the demand for high-quality products while also showcasing advancements in hardware and AI integration [3][5] Product Overview - Meta launched three models: Ray-Ban Meta Gen2, Oakley Meta Vanguard, and Meta Ray-Ban Display, each targeting different consumer segments [4] - The new models feature significant upgrades in pixel quality, stabilization, and battery life, with the Ray-Ban Meta achieving up to 30 hours of battery life with a charging case [5][6] - Pricing for the new models ranges from $379 for Ray-Ban Meta Gen2 to $799 for Meta Ray-Ban Display, which is competitive compared to similar products in the market [6] Technical Specifications - Meta Ray-Ban Display stands out with its display capabilities, featuring a resolution of 600x600, a field of view of 20 degrees, and a refresh rate of 90Hz, making it suitable for various environments [8] - The device includes a Meta Neural Band for hands-free control, allowing users to interact without needing to touch the glasses or their phones, although this feature has received mixed reviews [10] Market Dynamics - Despite the excitement around AI glasses, the industry faces challenges such as high return rates, with reports indicating return rates of 30% to 50% for various brands [21][22] - The current market is characterized by a lack of standardization among products, with various companies pursuing different technological paths, leading to a fragmented landscape [15][16] - The AI glasses market is experiencing rapid growth, with significant increases in sales volume, but this is coupled with high return rates, indicating consumer dissatisfaction [21][22] User Experience Challenges - Users have reported issues with comfort, functionality, and practicality, leading to a perception that current AI glasses are more of a novelty than a necessity [24][25] - The reliance on common chips for processing has limited advancements in weight, battery life, and overall functionality, posing a challenge for manufacturers [25][26] - Concerns regarding privacy and the complexity of after-sales service further complicate consumer acceptance of AI glasses [26]

Summary of Key Points from the Conference Call Industry Overview - The conference call focuses on the **Global Semiconductors, Hardware, Internet & Software** industry, specifically the **AI Server and Edge AI** market for **3Q25** [2][3]. Core Insights and Arguments 1. **Investment in Data Centers**: Total investment in upcoming and under-construction data centers is approximately **US$770 billion**. This includes significant projects like Oracle's multi-billion-dollar cloud contracts and a **4.5GW** investment for Stargate in the U.S. [3][27]. 2. **Capex Growth**: Consensus estimates for **2026 capex** from major cloud service providers (CSPs) have been raised by nearly **20%** compared to previous estimates, projecting total capex to grow at a **26% CAGR** from **2024-2027**, reaching around **US$500 billion** by 2027 [3][26]. 3. **Server Market Projections**: The global server market is expected to reach **US$450 billion** in 2026, with high-end GPU AI server shipments projected to grow at **45% CAGR** from **2024-2026** [4][39]. 4. **ASIC Adoption**: ASICs are projected to comprise nearly **40%** of total CoWoS-based AI chip shipments by **2026**, with increasing adoption among external customers [4][41]. 5. **High-End GPU Server Growth**: High-end GPU servers are expected to grow approximately **55%** in **2025** and **35%** in **2026** [4][39]. Financial Performance of Suppliers 1. **Supplier Performance**: The financial performance of suppliers in the AI supply chain is improving, with companies like Wiwynn and Accton projected to see revenue growth of **130%** and **111%** YoY respectively [43]. 2. **Broadcom's Orders**: Broadcom has received over **$10 billion** in new orders, indicating strong demand for ASICs [6][41]. Innovations in Edge AI 1. **AI Capabilities in Devices**: Innovations in Edge AI are evident across various devices, with Android phones and Apple products integrating more AI features. Gartner forecasts AI PC penetration to rise from **15%** last year to approximately **80%** by **2027** [7][28]. Company Ratings and Price Targets 1. **Chroma**: Rated **Outperform** with a price target of **NT$570** [11]. 2. **Delta**: Rated **Outperform** with a price target of **NT$630** [12]. 3. **Unimicron**: Rated **Outperform** with a price target of **NT$170** [13]. 4. **Quanta**: Rated **Underperform** with a price target of **NT$240** [14]. 5. **NVIDIA**: Rated **Outperform** with a price target of **$225** [20]. Important but Overlooked Content 1. **AI Funding Trends**: AI funding in the primary market remains strong, with **$40 billion** raised in **2Q25**, accounting for about **45%** of global startup funding [28]. 2. **Future Monitoring**: Key events to watch include capex guidance from CSPs and the progress of major projects like Stargate and Oracle's data center build-out [28]. This summary encapsulates the critical insights and projections from the conference call, highlighting the growth trajectory of the AI server market and the financial performance of key players in the industry.

Summary of Key Takeaways from the Conference Call Industry Overview - The report focuses on the Internet/e-Commerce industry, particularly large-cap Internet companies including Google, Meta, and Amazon, highlighting the impact of capital expenditures (capex) and depreciation costs on stock sentiment and performance [1][7]. Core Points and Arguments Positive Aspects - **AI as a Growth Driver**: AI is identified as a multi-dimensional growth driver for large-cap Internet companies, enhancing core businesses, creating new revenue streams, and improving operational efficiencies. For instance, Google is experiencing increased Search monetization and Cloud adoption due to rising demand for AI compute [2][25]. - **Capex Growth**: The sector's capex is expected to grow significantly, with estimates of 63% year-over-year growth in 2025 to $274 billion, followed by a 22% increase in 2026 to $333 billion. This growth is driven by investments in technical infrastructure, particularly for AI [7][10][14]. Risks - **Margin Pressure**: There is a risk of margin pressure due to a timing mismatch between scaling AI-driven revenue streams and the associated depreciation and amortization (D&A) expenses, which could impact profitability before new revenue uplifts materialize [3][61]. - **Shorter Lifespan of AI Infrastructure**: The rapid innovation cycles in AI may necessitate more frequent replacements of AI-specific assets, potentially leading to accelerated expense recognition and impacting earnings [3][65]. - **Capacity Overbuild**: The risk of overbuilding AI infrastructure could lead to supply exceeding demand, resulting in aggressive pricing strategies that may erode profitability [3][70]. Depreciation Expense Underestimation - The report suggests that the market may be underestimating the depreciation expenses for Google, Meta, and Amazon, particularly in 2026-2028, with significant divergences expected in 2027 [4][22]. Additional Important Insights - **AI Investments and Revenue Opportunities**: AI investments are expected to unlock new revenue opportunities beyond core businesses, such as subscription services for Google and advanced shopping capabilities for Amazon [50][54]. - **Operational Efficiencies**: Companies are likely to seek operational efficiencies to offset rising AI-related depreciation costs, potentially leading to lower headcount growth [57][58]. - **Revenue Growth vs. D&A Growth**: The combined revenue of Alphabet, Meta, and Amazon is projected to grow at 13% year-over-year in 2026, while combined D&A expenses are expected to increase by 33%, indicating potential margin compression [62][63]. Conclusion - The report emphasizes the critical role of AI in driving growth and efficiency for large-cap Internet companies while also highlighting significant risks related to margin pressure, asset lifespan, and potential overcapacity in the market. The anticipated growth in capex and the associated depreciation expenses will be key factors influencing the financial performance of these companies in the coming years [1][7][3].

Core Insights - Darui Electronics has identified Meta as a significant end customer, with ongoing collaborative projects that are expected to drive business growth in the coming year [1] - The company is actively advancing the research and application of composite materials in emerging AI hardware, including AI glasses and AR/VR devices [1] Company Developments - The collaboration with Meta is progressing smoothly, presenting new opportunities for business expansion in the next fiscal year [1] - The focus on composite materials indicates a strategic shift towards innovative technologies in the AI hardware sector [1]

企业资讯 市场回顾 上周五，美股三大指数收盘全线上涨，续创历史新高。截至收盘，道指上涨0.37%，报46315.27点；纳 指上涨0.72%，报22631.48点；标普500指数上涨0.49%，报6664.36点。全周来看，道指累计上涨 1.05%，纳指上涨2.21%，标普500指数上涨1.22%。 国际金价上周五反弹，结束两日连跌走势。截至9月19日收盘，COMEX黄金主力合约收报3719.4美元/ 盎司，上涨1.12%；伦敦现货黄金价格收报3684.65美元/盎司，上涨1.12%。上周，国际金价波动显著， 在美联储降息前创下历史新高，此后因投资者获利卖出回调。 国际宏观 据央视新闻，美国国会参议院19日否决众议院通过的一项临时拨款法案，推高部分联邦政府机构因资金 耗尽而"停摆"的风险。此前，美国众议院通过了为期七周的临时拨款法案，为政府提供资金直至11月21 日，避免即将到来的政府停摆。 据央视新闻，美国总统特朗普19日签署公告，将企业为H-1B签证申请人支付的费用提高至10万美元， 并称此举旨在确保美国引进的是高技能、不能被美国人替代的人才。此前，企业为H-1B签证支付的费 用通常在数千美金。 据彭博社 ...

Core Viewpoint - Meta has launched its first smart glasses with a display, the Meta Ray-Ban Display, and a new version of Ray-Ban Meta (Gen 2) smart glasses, along with the Hyperscape feature for Quest 3 and Quest 3S VR headset users, indicating a significant advancement in wearable technology and AI integration [1][2][3]. Product Launches - The Meta Ray-Ban Display features a full-color high-resolution screen on the right lens, capable of displaying key information without obstructing the user's view, supporting functions like subtitle translation, navigation maps, and camera previews, with a mixed-use battery life of approximately 6 hours and a total of 30 hours with a charging case [2][3]. - The new Ray-Ban Meta (Gen 2) smart glasses offer an extended battery life of up to 8 hours, nearly double that of the first generation, and include a 12-megapixel ultra-wide camera capable of 3K video recording at 60 frames per second, with an upgraded charging case providing an additional 48 hours of power [2][3]. Hyperscape Feature - The Hyperscape feature allows users of the Quest 3 and Quest 3S VR headsets to scan their real-world environment and convert it into a virtual space, enhancing the immersive experience [3]. Industry Insights - The development of smart glasses and mixed reality (MR) ecosystems is maturing, with advancements in AI technology expected to drive sales growth [1][4]. - The company has multiple chips integrated into the Meta Ray-Ban smart glasses and Quest MR series, which are expected to enhance performance and user experience in next-generation wearable and immersive devices [3]. Financial Projections - The company maintains its performance forecast, expecting net profits attributable to the parent company to reach 393 million, 523 million, and 621 million yuan for the years 2025-2027, corresponding to price-to-earnings ratios of 51, 39, and 33 times based on the stock price as of September 19, 2025 [1][4].