Meta的交易将Scale AI的估值从去年的138亿美元推高至280亿美元，凸显了扎克伯格在AI军备竞赛中的焦虑与野心。对 于Meta而言，收购Scale AI不仅能获得关键的数据标注能力，更重要的是获得Wang及其团队对竞争对手AI模型构建方式 的深度了解。 硬·AI 作者 | 赵雨荷 值得注意的是，Meta选择收购49%股份而非全资收购，这一结构设计显然旨在规避更严格的监管审查。 目前Meta正因收购Instagram和WhatsApp面临反垄断调查，全资收购可能引发监管部门审查。 这笔交易凸显了扎克伯格在AI军备竞赛中的焦虑与野心，也反映了当前AI领域的激烈竞争态势。谷歌此前 投资Character.AI，微软重金押注Inflection AI，科技巨头们都在通过大额投资而非直接收购的方式，获得 顶尖AI人才和技术。 01 Scale AI： AI训练数据"独角兽" Scale AI成立于2016年，专注于提供大规模数据标注服务——也就是为ChatGPT等高级AI工具提供经过精 心整理的训练数据。 该公司去年估值为138亿美元， 2024年营收约8.7亿美元，预计今年将突破20亿美元大关。但公司仍处于 ...

Core Viewpoint - Apple is enhancing its software ecosystem with a new Foundation Models framework that allows third-party developers to integrate built-in large language models (LLM) into their apps, alongside a significant redesign of its operating systems and various applications [1][6][9]. Group 1: Foundation Models Framework - The new Foundation Models framework enables third-party developers to utilize Apple's LLM directly in their applications, enhancing the development capabilities based on Apple Intelligence [5][6]. - This framework is currently in testing and will be available for public beta next month, featuring built-in functionalities like guided generation and tool invocation [8]. Group 2: Operating System Naming and Design - Apple has simplified its operating system naming convention to a yearly format, with the latest versions named 26, such as iOS 26, watchOS 26, and macOS Tahoe [3][9]. - The new design, referred to as "Liquid Glass," will be applied across various devices, providing a semi-transparent and adaptive user interface that enhances user experience [9][11]. Group 3: iOS 26 Features - iOS 26 introduces a redesigned camera app with a simplified menu and a more fluid user interface, along with updates to the Photos app [14]. - The Phone app now integrates voicemail summaries and call screening features, while FaceTime has been revamped to enhance user interaction [16][18]. - The Messages app has added group chat polling and cash transfer functionalities, improving communication and transaction capabilities [20][21]. Group 4: Additional Features and Applications - New features in Apple Music include lyric translation and an auto-mix function, enhancing the music experience for users [25][27]. - The Maps app now includes preferred routes and visited places functionalities, while the Wallet app expands to support digital IDs and enhanced tracking of orders [29][31]. - A new Games app will consolidate all mobile gaming functions, allowing users to easily access and manage their games [34][35]. Group 5: watchOS 26 and macOS Tahoe - watchOS 26 will feature AI-enhanced functionalities like Workout Buddy, which provides real-time insights and encouragement during workouts [42]. - macOS Tahoe will adopt the Liquid Glass design and introduce features like Live Activity, allowing seamless integration of activities initiated on iPhone [46][48]. Group 6: visionOS 26 and iPadOS 26 - visionOS 26 will include customizable widgets and improved spatial features, while iPadOS 26 will enhance multitasking capabilities and introduce a new window system [51][52].

北京生数科技有限公司首席执行官骆怡航发表了主题演讲——"多模态生成：从模型走向生产"，主要围绕 多模态大模型，特别是视频生成在产业落地中的机遇、挑战，并分享了生数科技（Vidu）的解决方案和成 果。 以下是演讲亮点： 多模态大模型迎来规模化生产落地的拐点：第一，我们看到技术迭代非常迅速，音视频的生成模型无论在 效果、速度、成本上都快速提升。其次，行业需求特别旺盛。第三，很多行业视频内容相关的各种产业落 地节奏加快。 今年包括再往后要同时具备四个条件：内容的创意，内容质量、生成的效率和生产的成本。 如果具备了内容的质量好于传统方式，同时生产效率和生产成本，在我看来效率必须比传统的方式要至少 百倍的提升。 对于生数科技来讲，我们聚焦在多模态生成，现在主要以视频生成为主，包括音视频的部分，未来我们会 延展到3D叙事空间等等。目前我们聚焦在专业的用户和企业用户，致力于把模型推动到8大行业、30大场景 里面。 Vidu 2.0把速度极大做了提升，可以达到 5 秒技术生成。同时Vidu Q1 进一步提升，包括高清的版本，还有 首尾帧，还有动漫等方向。同时对于音效还有音频我们做了深化。 从Vidu上线以来，专业创作的占比增 ...

Core Viewpoint - Broadcom's stock price has surged 70% over the past two months, reflecting investor optimism regarding its growth in the AI chip sector, with a market capitalization exceeding $1 trillion, making it the seventh-largest publicly traded company in the U.S. [2] Group 1: Financial Performance and Valuation - Broadcom's expected P/E ratio is 33.6, which is a 23% premium over the Philadelphia Semiconductor Index and the highest premium in nearly a decade, as well as a 13% premium over Nvidia [1][8] - The company anticipates AI processing and networking chip revenue to reach $5.1 billion in Q3, representing a 60% year-over-year increase and accounting for about one-third of total sales [2] Group 2: Customer Dependency and Risks - Broadcom's AI revenue is highly dependent on Google, with TPU orders constituting over 80% of its AI revenue, posing significant risks if Google alters its strategy or shifts to other suppliers [6][4] - The company is expected to diversify its AI customer base, but this transition may not be evident until the second half of the 2026 fiscal year [6] Group 3: Market Challenges - The custom AI chip market presents challenges, including performance issues where Google's latest custom AI chip performs at only about half the capability of Nvidia's corresponding products [10] - Broadcom faces increasing competition not only from Nvidia but also from other chip manufacturers like Marvell and MediaTek, which are gradually encroaching on its market share [11] Group 4: Long-term Growth Potential - Despite the challenges, analysts remain optimistic about Broadcom's growth prospects, suggesting that the overall rise in AI demand will support its revenue [13] - Broadcom's target price has been raised from $240 to $300, with expectations of significant earnings growth driven by AI sales, projecting around $10 billion in earnings per share by fiscal year 2027 [14]

Core Insights - The core insight of the article is that the adoption rate of AI among U.S. enterprises has significantly increased, indicating a robust trend in AI investment, particularly benefiting the semiconductor industry [2][3][5]. Group 1: AI Adoption Rates - The AI adoption rate among U.S. enterprises rose from 7.4% in Q4 of the previous year to 9.2% in Q2 of this year, with large enterprises (250+ employees) showing a higher adoption rate of 14.9% [2][9]. - Industries such as education, information, finance, and professional services reported the largest increases in AI adoption, with growth exceeding 3 percentage points [9]. - Medium-sized enterprises (100-249 employees) are expected to see the largest increase in AI adoption over the next six months, with an anticipated rise of 4.7 percentage points to 14.6% [9]. Group 2: Semiconductor Industry Impact - The semiconductor industry is projected to benefit significantly from the AI investment wave, with revenue forecasts for 2025 being raised by $200 billion due to the AI boom [5]. - Analysts have also increased revenue forecasts for other AI hardware companies by $105 billion since the launch of ChatGPT [5]. - The semiconductor industry's revenue is expected to grow by 36% by the end of 2026 compared to current levels, reflecting the sustained interest in AI investments [2][5]. Group 3: Broader Economic Implications - Despite the rapid growth in AI adoption, the impact on the labor market remains limited, with most labor market indicators showing no significant effects [10]. - AI-related job vacancies currently account for 24% of all IT job vacancies and 1.5% of all job postings [11]. - The productivity paradox suggests that while AI adoption is limited, the effects on productivity in deployed areas are significant, with average productivity increases of 23% in academic research and approximately 29% in corporate examples [11][12].

Core Viewpoint - Perplexity is experiencing significant growth, with a monthly query volume reaching 780 million and a potential to become a major player in the AI search engine market if the current growth rate of over 20% is maintained [1][2][4]. Group 1: Growth Metrics - The CEO of Perplexity, Aravind Srinivas, announced that the company processed 780 million queries in May, marking a growth of over 20% month-over-month [2]. - Compared to its early days in 2022, when the daily query volume was only 3,000, it has surged to 30 million queries per day [3]. Group 2: Comet Browser Development - Perplexity is developing the Comet browser, which is positioned as a key growth engine for the company, promising "infinite retention" advantages [5][6]. - The core idea behind Comet is to shift the role of AI from merely providing answers to actively executing tasks on behalf of users, aiming to streamline the browsing experience [6][7]. Group 3: Strategic Vision and Challenges - Srinivas describes Comet as a "cognitive operating system," intending to fundamentally change how users interact with the internet by providing seamless assistance throughout their browsing sessions [7]. - The strategy mirrors Google's early growth path, focusing on deep user data mining to build a precise advertising ecosystem, although the exact launch date for Comet has not been disclosed [8]. - The challenges ahead include technological implementation difficulties, user habit resistance, and competition from established giants like Google and Microsoft [8].

Core Viewpoint - The collaboration between CoreWeave, NVIDIA, and IBM has achieved the largest scale MLPerf Training v5.0 test result in history, utilizing 2496 GB200 Grace Blackwell chips, showcasing the capabilities of CoreWeave's AI-optimized cloud platform [2][3] Group 1: MLPerf Training v5.0 Test Results - The GB200 NVL72 cluster completed the training of the largest and most complex Llama 3.1 405B model in just 27.3 minutes, achieving over two times the training performance compared to similar scale clusters [3] - This test result highlights the significant performance leap provided by the GB200 NVL72 architecture and CoreWeave's strong infrastructure for consistent and high-performance AI workloads [3] Group 2: Industry Participation and Growth - The MLPerf Training v5.0 test received a record total of 201 performance test submissions from 20 organizations, indicating a new high in industry participation [6] - The introduction of the new Llama 3.1 405B large language model (LLM) pre-training benchmark has replaced the previous GPT-3 based benchmarks, reflecting the growing importance of large-scale training [5][6] - The participation of new organizations such as AMD, IBM, MangoBoost, Nebius, and SCITIX in the MLPerf Training tests has been particularly welcomed, emphasizing the increasing focus on energy efficiency in AI training systems [6]

据报道，亚马逊即将在其旧金山办公室的"人形公园"内测试人形机器人，以取代部分人工配送岗位，削减运营成本；公司 还同时在测试Rivian电动货车与机器人的互动，为其主营业务——全球包裹配送——的自动化铺平道路。 硬·AI 作者 | 李笑寅 编辑 | 硬 AI 亚马逊正在悄然布局自动驾驶+人形机器人配送，有望颠覆物流行业格局。 据媒体援引知情人士消息， 亚马逊即将在其旧金山办公室的"人形公园"内测试人形机器人，目标是取代部 分人工配送岗位，削减巨额运营成本。 亚马逊推动配送自动化的动机显而易见：成本控制。 作为全球最大的电商平台之一，亚马逊雇佣数十万名 配送工人，人工成本高企。 而人形机器人交付仅是第一步，要完全取代送货员，还需要实现他们驾驶车辆的自动化。如果人形机器人 能够与自动驾驶车辆结合，甚至在人类司机旁协助配送，整体效率将显著提升。 此举也将为其主营业务——全球包裹递送——的自动化铺平道路。 媒体还透露，亚马逊已将一辆Rivian电动货车放置在这个"人形公园"内，以帮助测试机器人如何与车辆互 动。 这一内部项目模仿了自动驾驶汽车开发的路径——首先在封闭环境测试，然后才会扩展到公共领域。 据报道介绍，这个相 ...

博通周二宣布开始出货其最新的数据中心交换机芯片Tomahawk 6，声称其以太网交换容量是目前市场上任何以太网交换 机带宽的两倍。这款芯片专为运行AI处理器集群的数据中心需求而设计。消息后，博通股价应声上涨近3.3%，市值首次 逼近1.2万亿美元。 硬·AI 作者 | 鲍亦龙 编辑 | 硬 AI 博通在财报前夕突然抛出"突破性"数据中心交换机芯片，股价创历史新高。 6月3日，当所有人都在等待博通财报时，这家芯片巨头却宣布开始出货其最新的数据中心交换机芯片 Tomahawk 6，声称 其以太网交换容量达到每秒102.4太比特，是目前市场上任何以太网交换机带宽的两 倍 。 这款芯片专为运行AI处理器集群的数据中心需求而设计 。 博通核心交换业务高级副总裁兼总经理Ram Velaga在新闻稿中强调： Tomahawk 6不仅仅是升级——它是一个突破。它标志着AI基础设施设计的转折点，将最高带宽、 功率效率和自适应路由功能结合到一个平台中。 据媒体报道，该芯片采用台积电3纳米工艺制造，相比前代Tomahawk 5的5纳米工艺实现了制程升级。博 通强调， 带宽翻倍的直接好处是减少了系统所需的网络设备数量，从而降低功耗并 ...

周二，英伟达股价上涨约3%至141.22美元，市值达到3.45万亿美元，超越微软的3.44万亿美元。目前，英伟达的预期市 盈率约为29倍，远低于其历史均值，PEG指标也低于Mag 7中其他成员的水平，意味着股价更被低估。 在特朗普贸易政策威胁和市场波动加剧之际， 周二，英伟达股价上涨约3%至141.22美元，市值达到3.45 万亿美元，超越微软的3.44万亿美元，1月以来首次重登全球市值最高公司宝座。 这意味着英伟达的市值在两个月内反弹了1万亿美元。去年6月以来，英伟达一直与苹果和微软在市值榜首 位置交替，上一次登顶还是在今年1月24日。 英伟达的强劲表现也带动了整个芯片板块走高。Broadcom上涨3%，美光科技上涨4%，追踪一篮子芯片 股的VanEck半导体ETF上涨2%。 编辑 | 硬 AI 尽管面临关税担忧，但英伟达的股价在过去一个月飙升近24%，自4月低点以来上涨逾45%，证明了市场 对其核心业务和增长前景仍信心十足。 AI狂潮仍未降温！英伟达市值超越微软，再度加冕全球最贵公司。 Globalt Investments高级投资组合经理Thomas Martin表示： "那些关于英伟达的问题已经得到了 ...