Core Insights - The article discusses NVIDIA's groundbreaking research on the Agentic Variation Operator (AVO), which represents a significant advancement in autonomous evolutionary search in software development [1][2][5]. Research Overview - AVO is a new type of evolutionary variation operator that replaces traditional fixed mutation and crossover methods with autonomous coding agents, achieving impressive results in optimization tasks [2][7]. - The research highlights the limitations of existing frameworks that confine large language models (LLMs) to generating candidate solutions without the ability to actively consult references or test modifications [5][10]. Performance Achievements - AVO was applied to optimize the Multi-head Attention (MHA) kernel on NVIDIA's Blackwell B200 GPU, achieving a throughput of 1668 TFLOPS, surpassing cuDNN by 3.5% and FlashAttention-4 by 10.5% [9][19]. - Over a continuous 7-day period, the AVO agent explored more than 500 optimization directions and produced 40 effective iterations without human intervention [16][19]. Methodological Innovations - AVO transforms the role of AI from a passive code generator to an active evolutionary agent capable of self-directed exploration and iterative improvement [10][12]. - The agent can autonomously access previous code versions, domain-specific knowledge bases, and evaluation tools, allowing for continuous enhancement over extended periods [7][10]. Generalization and Adaptability - The optimizations discovered by AVO are not limited to specific scenarios; they can be effectively transferred to different tasks, such as adapting the optimized MHA kernel to Grouped-query Attention (GQA) with only 30 minutes of additional self-adaptation [18][23]. - The research indicates that the optimizations identified by the agent reflect genuine hardware-level reasoning rather than superficial code changes, demonstrating the agent's capability to handle complex hardware interactions [24][25].

Group 1: Stock Market Movements - China Merchants Energy (01138) fell over 5%, currently down 4.09%, as the company assesses safety conditions regarding its vessels in the Strait of Hormuz [1] - Shenzhen Expressway (00548) dropped over 9%, currently down 7.86%, after reporting 2025 revenue of approximately 9.264 billion yuan, a 0.2% year-on-year increase, and a net profit of about 1.149 billion yuan, up 0.38% [1] - Television Broadcasts (00511) rose nearly 3% after announcing a 2025 revenue of 3.192 billion HKD, a 2.02% decrease, and a shareholder loss of 491 million HKD [2] - Chaoyun Group (06601) increased over 6%, currently up 4.07%, reporting a revenue of 1.988 billion yuan, a 9.24% year-on-year growth, and a net profit of 224 million yuan, up 9.98% [2] - Golden stocks fell again, with companies like Zijing Mining (02899) and Shandong Gold (01787) experiencing declines of 3.09% and 4.02% respectively, as spot gold prices dipped below 4500 USD [3] Group 2: Company Earnings and Forecasts - GSK Group (01285) surged over 40%, currently up 31.87%, after announcing an expected net profit of between 101.4 million and 111.1 million yuan for the year ending December 31, 2025, representing a year-on-year increase of approximately 210% to 230% [4] - Aoda Holdings (09929) fell over 12%, currently down 16.19%, reporting a revenue of 66.258 million HKD, a 27.75% decrease, and a net loss of 18.856 million HKD [5] - Peijia Medical (09996) rose nearly 4%, reporting a revenue of 713 million yuan, a 15.82% year-on-year increase, and a narrowed net loss [3] Group 3: Market Trends and Sector Performance - The pork concept stocks collectively declined, with COFCO Joycome (01610) down 3.57% as the average trading price of lean pigs fell below 10 yuan/kg, marking a 15-year low [3] - Semiconductor stocks saw a strong start, with Arm Holdings (ARM.US) rising 16.38% after announcing its entry into chip manufacturing, aiming for 15 billion USD in annual revenue within five years [8] - AI application software stocks collectively rose, with Braze (BRZE.US) increasing by 19.87% [8]

Core Viewpoint - Nvidia's stock has the potential to increase by 150% in the coming years, contingent on sustained growth in artificial intelligence (AI) infrastructure spending [1] Group 1: AI Infrastructure Spending - Data center AI spending is expected to grow rapidly, with predictions of reaching $1.4 trillion by 2030, up from approximately $500 billion last year [2] - The majority of this spending will focus on computing power, particularly Nvidia's GPUs, while networking is anticipated to grow even faster [2] Group 2: Market Share and Competition - Nvidia must maintain its market-share lead in AI chips, as competitors like Advanced Micro Devices are expected to gain some market share through partnerships with OpenAI and Meta Platforms [3] - Despite competition, Nvidia is well-positioned to remain the market leader due to the ecosystem it has developed around its chips [3] Group 3: Business Transformation - Nvidia has evolved from a GPU manufacturer to a comprehensive AI infrastructure solutions company, with its networking business being the fastest-growing segment [5] - The company has expanded into other chip markets, including CPUs and data processing units (DPUs), and has strengthened its software capabilities [5] Group 4: Strategic Acquisitions - Nvidia's acquisitions of Groq and SchedMD have enhanced its position for the future of inference and agentic AI, integrating language processing units (LPUs) and developing solutions for AI agents [6] - The new Vera CPU has been specifically designed for agentic AI, indicating Nvidia's commitment to evolving its product offerings [6] Group 5: Revenue Projections - Nvidia has diversified its revenue streams beyond GPUs, with projections indicating that China could contribute an additional $25 billion annually [7] - This growth could lead to earnings per share (EPS) of $20 or more by fiscal 2030, potentially driving the stock price to $450 by the end of 2028 [7]

马斯克计划设立半导体工厂Terafab，力争每年生产按消耗电力换算相当于1太瓦规模的AI半 导体。据他表示这是当前全球供应能力的50倍。有估算认为，生产1太瓦半导体需要5~13万 亿美元的设备投资，金额过于庞大…… 美国企业家埃隆·马斯克正面向量产人工智能（AI）半导体，展开继电动汽车和航天业务之后 的第三次押注。将建设堪称"美国版台积电"的先进工厂，用于处理AI的庞大数据。通过特斯 拉和SpaceX颠覆产品制造常识的马斯克，此次将面临估算额高达13万亿美元的巨额费用。 马斯克提出的构想是，作为其执掌的特斯拉与SpaceX的共同业务，在美国南部得克萨斯州奥 斯汀设立半导体工厂"Terafab"。 马斯克于2月将AI开发公司xAI并入SpaceX，还让开发AI机器人的特斯拉向SpaceX出资，推 进了旗下企业的联盟化。马斯卡将AI视为整个企业群的增长战略，为了确保各公司使用的半导 体，将启动自主进行量产的宏大项目。 | 马斯克以Al 为核心整合企业群 | | | --- | --- | | 特斯拉 | 自动驾驶EV、人形机器人"擎天柱u" AI半导体开发 | | SpaceX | 人造卫星与太空数据中心 | | ...

Core Viewpoint - Reflection AI, a startup supported by Nvidia, is negotiating to raise $2.5 billion at a valuation of $25 billion, reflecting its pre-funding valuation [1] Group 1: Company Overview - Reflection AI is in discussions to secure $2.5 billion in funding [1] - The proposed valuation of $25 billion is based on the company's pre-investment worth [1] Group 2: Investment Details - JPMorgan is considering participation in this funding round through its Security and Resiliency Initiative [1]

机器之心编辑部 这应该是今天刚刚出炉的、最炸裂的文章。 在很多算子开发的微信群组，已经掀起了轩然大波。 「这或许是超人类智能在软件领域的真正首次展露。」英伟达许冰刚刚在 X 上发出了如此断言。他所评论的，正是他与 Terry Chen 和 Zhifan Ye 为共同一作的一项 英伟达新研究 AVO 。 黄之鹏的 X 推文 有意思的是，在 X 推文中，许冰还分享说一年半之前他与 Terry Chen 刚开始在英伟达研究智能体编程时，他们还不懂 GPU 编程，「所以从一开始我们就致力于开 发完全自动化、无需人工干预的系统。」他们称之为「 盲编程（blind coding） 」。 在本周四刚刚提交到 arXiv 上的这项研究中，英伟达构建了 Agentic Variation Operator（AVO） ，这是一类新型进化变异算子，它用自主编码智能体取代了经典 进化搜索中固定的变异、交叉和人工设计的启发式方法，并取得了相当震撼的实际表现。 许冰表示：「在一些经过高度优化的注意力机制工作负载中，智能体在没有人工干预的情况下，即可在优化循环中连续搜索 7 天， 从而超越几乎所有人类 GPU 专 家 。」——AVO 的如 ...

Core Insights - Nvidia has been the leader in the AI hardware market, but competition is increasing as companies develop their own hardware solutions to reduce reliance on Nvidia [1] - Alphabet's TPU is a direct competitor to Nvidia's GPUs, and OpenAI is also seeking alternatives, finding a partner in Broadcom [2][4] Company Developments - Broadcom collaborates with various companies, including Google, Microsoft, Amazon, and Meta, to develop custom chips tailored for specific AI applications [3] - OpenAI and Broadcom entered a multiyear partnership to co-develop 10 gigawatts of custom AI accelerators, indicating a shift towards specialized hardware [4] - Anthropic, OpenAI's competitor, is also moving towards custom chips, further highlighting the trend away from general-purpose GPUs [5][6] Financial Performance - Broadcom's revenue is projected to increase by 24% in 2025, reaching $63.8 billion, with diluted earnings per share (EPS) growing by 40% [8][9] - The company has a net profit margin of 36.57% and a debt-to-equity ratio of 0.83, indicating a strong financial position [9] - Broadcom anticipates its AI semiconductor revenue to double to $8.2 billion this year, driven by partnerships like the one with OpenAI [9]

周三（3月25日）的美股市场，完全被地缘政治的拉锯战与原油市场的剧烈波动所主导。美国向伊朗提交15点停火方案的消息一度引发盘前狂欢，但随着 伊朗态度趋于强硬，市场情绪迅速回归现实。 最终，美股三大指数在震荡中收涨，但表面普涨的背后，美股与美债、原油的传统联动关系出现了罕见的阶段性背离。透过盘面波动，看清主力资金在关 键点位的真实意图，比单纯关注收盘数字更为重要。 （美股债油日内走势对比） 地缘博弈下的原油"深V 原油市场是今天多空博弈的主战场。早盘阶段，受美国Z府通过中间方递交"15点停火方案"并提议停火一个月的消息刺激，布油一度急速下挫3%。但这股 乐观情绪转瞬即逝，伊朗方面迅速拒绝了该方案，明确表示不允许特朗普决定停战时间表，并强调现阶段的信息交流"并非谈判"。 随着和谈预期降温，油价走出了极致的V型反转。美油收复日内全部跌幅并最终涨至91美元上方，布油突破103美元。不仅原油走强，美国国内精炼油品 价格也持续高企，柴油价格已连续九天站稳每加仑5美元的关口。 | 1 Futures & Indexes | Last | Change | % Change | Last Updated | | --- | ...

Group 1 - Nvidia-backed startup Reflection AI is in discussions to raise $2.5 billion at a valuation of $25 billion, indicating strong investor interest in AI technologies [1][2] - The proposed valuation of $25 billion represents a pre-money valuation, which is an increase from the previously targeted valuation of over $20 billion [2][3] - JPMorgan Chase is considering participating in this funding round through its Security and Resiliency Initiative, highlighting the strategic interest from major financial institutions [2]

2026年3月下旬，科技圈的热度褪去了年初的盲目狂欢，进入"技术落地与争议并存"的关键周期。AI智能体从全民"养虾"热潮中冷静下来，脑机接口、核 聚变等硬核科技迎来里程碑突破，英伟达GTC大会重新定义AI产业底层逻辑，国产科技力量持续打破垄断。六大热点交织之下，科技产业正从"概念炒 作"向"价值落地"加速切换，每一个突破与争议，都在勾勒未来科技的清晰轮廓。 AI智能体：从"全民狂欢"到"理性回归"，落地痛点凸显 如果说3月上旬的科技圈被AI智能体的热潮席卷，那么下旬的关键词则是"降温与反思"。OpenClaw（俗称"龙虾"）的热度虽未完全消退，GitHub日下载量 仍维持在10万+级别，但"花钱卸载"的用户反馈逐渐增多，暴露了当前AI智能体的核心短板——稳定性不足、实用性脱节。 这场热潮背后，是大厂的竞速博弈：字节ArkClaw、腾讯WorkBuddy、阿里CoPaw、智谱AutoClaw等产品密集上线，试图抢占AI下半场的"智能体赛道"。 但正如虎嗅智库在近期AI闭门会上所观察到的，很多AI项目本质上仍停留在"玩具"阶段，即便落地到ToB领域，也难以形成真正的业务价值闭环。多位企 业技术负责人坦言，AI智能 ...