截至2026年1月20日 10:03，中证港股通医药卫生综合指数(930965)成分股方面涨跌互现，时代天使领涨 4.83%，映恩生物-B上涨1.86%，联邦制药上涨1.23%；药明康德领跌。港股医药ETF(159718)最新报价 0.96元。 消息面上，医药魔方NextPharma数据库显示，截至2025年12月31日，中国创新药商务拓展（BD）出海 授权全年交易总金额达到1356.55亿美元，首付款70亿美元，交易总数量达到157笔，远超2024年全年 519亿美元和94笔，各个维度的数据统计均达到了历史新高。 华鑫证券表示，2025年中国创新药BD出海授权交易总金额达1356.55亿美元，首付款70亿美元，数量达 157起，均创历史新高；进入2026年开局延续强劲势头，荣昌生物PD-1/VEGF双抗RC148获艾伯维6.5亿 美元首付款授权，宜联生物与罗氏就B7H3靶向ADC达成合作。此外，葛兰素史克Bepirovirsen治疗慢性 乙肝III期研究取得积极结果，Arrowhead公布RNAi疗法在减重领域的新进展，小核酸药物临床价值持续 获得验证。国内头部企业如中国生物制药、恒瑞医药、石药集团等正加速 ...

这一里程碑式的进展正值人工智能硬件投资蓬勃发展之际。Groq Inc.去年在与英伟达达成战略合作之 前，完成了一轮7.5亿美元的融资。其他近期交易还包括Cerebras Systems Inc.以220亿美元的估值寻求10 亿美元的新融资，以及Etched完成的5亿美元融资。 知情人士透露，募集资金将用于FuriosaAI第二代RNGD芯片的量产、全球业务扩张以及第三代芯片的研 发。他们还表示，这家初创公司计划最早于2027年上市。其中一位知情人士称，FuriosaAI将于本月晚 些时候从台积电(TSM.US)收到首批量产的RNGD人工智能芯片。 FuriosaAI由曾任职于三星电子和AMD的资深人士June Paik于2017年创立，专注于高效人工智能推理芯 片的研发。据该公司称，其RNGD芯片(发音为"renegade")的每瓦推理性能是传统图形处理器的2.25倍。 去年7月，FuriosaAI拿下首个大客户LG。 据知情人士透露，人工智能芯片设计公司FuriosaAI计划在首次公开募股前通过一轮融资筹集高达 5 亿 美元，以挑战行业巨头英伟达(NVDA.US)。据知情人士透露，这家总部位于首尔的初创公司 ...

公众号记得加星标⭐️，第一时间看推送不会错过。 NVIDIA正在强化其数据中心布局，并推进个人和边缘AI计算平台的发展，规划了从N1和N1X到下 一代N2和N2X系列的AI时代演进路线。据供应链企业透露，基于NVIDIA最新的技术蓝图，采用 N1X架构的Windows on Arm (WoA)平台NB型号预计将于2026年第一季度发布，初期目标市场为消 费级；其他三个版本将于第二季度上市；而下一代N2系列预计将于2027年第三季度接棒。 NVIDIA正凭借其AI优势强势进军PC市场，给x86架构的英特尔和AMD两大巨头施加压力。目前， NVIDIA仅发布了搭载N1X处理器的DGX Spark，这款产品号称是全球最小的AI台式电脑(DT) 。搭 载N1X/N1平台的笔记本电脑原计划于2025年中期上市，但在今年1月初的CES 2026上仍未亮相。 近年来，NVIDIA积极拓展AI PC市场份额，推出了面向台式机和笔记本电脑的N1/N1X系列处理 器。这标志着NVIDIA在WoA阵营中重获领先地位，挑战英特尔和AMD在PC领域的统治地位。 供应链消息人士指出，DGX Spark 采用桌面式设计，可提供 1 pet ...

Core Viewpoint - 2026 is anticipated to be a significant year for certain stocks, particularly Nvidia, Nebius Group, and The Trade Desk, which are expected to show strong performance as fourth-quarter results from 2025 are released [1][2]. Nvidia - Nvidia is the largest company by market cap, driven by its dominant position in graphics processing units (GPUs) for artificial intelligence (AI) computing [3]. - The company has reported being sold out of cloud GPUs, indicating sustained demand for its products [4]. - Nvidia's fiscal fourth quarter 2026 earnings are set to be reported on February 25, with expectations that its fiscal 2027 guidance will exceed current market expectations [6]. Nebius Group - Nebius Group is gaining attention for its services that connect high-end Nvidia GPUs to computing clusters, experiencing significant demand [7]. - The company reported a remarkable 355% year-over-year growth in the third quarter, with an annual run rate (ARR) of $551 million [8]. - Nebius has dramatically increased its 2026 revenue projection, expecting an ARR of $7 billion to $9 billion by the end of 2026, which could lead to rapid stock appreciation if achieved [8][10]. The Trade Desk - The Trade Desk faced challenges in 2025, being one of the worst performers in the S&P 500 due to issues with its AI-powered ad buying platform and lack of political advertising revenue [11]. - In 2026, the company is expected to resolve its platform issues and avoid difficult year-over-year comparisons, potentially leading to a significant stock increase [12]. - Currently trading at less than 18 times forward earnings, The Trade Desk presents a value opportunity compared to the S&P 500 average of 22.4 times [12][14].

AIDC 电源革命开启，2026 从预期到现实 20260119 摘要 数据中心芯片功耗和机柜功率密度显著提升，英伟达 Rubin Ultra 芯片 开伯尔机柜功率已达兆瓦级别，预计北美 AI 数据中心到 2028 年新增功 率达 70 吉瓦，全球总新增功率预计达百吉瓦级别，对供电系统提出更 高要求。 英伟达和谷歌均提出数据中心供电四步走战略，最终目标一致，即采用 中压整流器或固态变压器（SST）将中压交流一步降至 800 伏直流，提 高转化效率，OCP 组织也在不断迭代数据中心供电标准，提升整体效率 与性能。 一次电源（柜外电源）未来发展方向是高压直流（HVDC）和固态变压 器（SST），SST 难点在于高频变压器技术，目前仅有少数企业交付 SST 产品或样机，HVDC 技术难点在于高功率 PSU 模块。 二次电源（柜内电源）发展趋势是单颗 PSU 功率不断提升，需采用碳化 硅和氮化镓等第三代半导体及新型拓扑结构，同时，从 HVDC 或 SST 主 机输出的 800 伏直流需通过 DCDC 转换器降至服务器可用电压。 AI 芯片电容市场由三星（原材料来自博迁新材）和村田主导，随着 AI 技 术发展，对高端电 ...

英伟达+特斯拉 FSD 深度体验交流 20260119 摘要 Waymo 在软件应用和响应速度方面表现出色，提供完善的综合体验， 但其基于规则的系统在限定区域内表现最佳，可扩展性受限，且面临技 术栈转换为端到端模式的挑战。 特斯拉 Robotaxi 不依赖高清地图，使用开源地图数据，覆盖更多路段， 提供更完整的末端体验，且价格较低。但投放车辆数量较少，用户等待 时间较长，私密性不如 Waymo。 特斯拉 Robotaxi 的运行模型与量产车上的 FSD V14 模型基本一致，更 注重泛化性，但软件稳定性和极低故障率是关键挑战，需要极高的系统 稳定性。 英伟达展示了搭载奔驰 CLA 车型的端到端自动驾驶模型，表现超出预期， 计划 2026 年第一季度覆盖整个加州，并逐步扩展至整个北美市场，但 短期内仍需验证大规模车队的数据采集及稳定运行能力。 地勤服务对于自动驾驶汽车至关重要，特斯拉通过设计电动门来减少地 勤需求，但完全无人驾驶仍需地勤支持。 Waymo 是目前全球范围内 Robotaxi 运营业务规模最大的一家公司，拥有 2,500 辆车。尽管这一数字与人们预期的上百万辆或几十万辆有明显差距，但 在实际部署到 ...

近期，谷歌的 Nested Learning 引发了一场模型界的记忆地震。 很多人重新意识到，大模型不必永远是「训练完就封存」的只读权重，它也可以在推理过程中继续变化。在 Nested Learning 里，当模型 读到新的上下文时，它不只是把文本塞进注意力的缓存里临时翻找，而是允许自己在推理过程中更改参数，让新信息变成它内部记忆的 一部分。 但就在人们还在消化这个想法时，英伟达在2025年12月28日给出了一个更激进的答案，一篇名为《End-to-End Test-Time Training for Long Context》的论文。谷歌的记忆增强路线，还在努力解决记忆问题，把过去重要的东西保存得更完整。但英伟达的研究人员则认为，记忆 其实就是学习，「记住」就是「继续训练」。 如果沿着时间线往回翻，你会发现 TTT（test-time training）并不是凭空出现的发明。 早在2013年，Mikolov 等人就在语言模型里尝试过 dynamic evaluation。当时放的是让模型解除冻结，在测试文本上继续用下一词预测的 交叉熵损失 CE（也就是我们最经常理解的大语言模型的参数学习损失目标）做小步梯度 ...

2026年1月 20 星期二 信息来源：网络 编辑-楓華 利率预期：强劲的就业市场直接锁定了利率走向。目前市场几乎确信美联储将在 1 月 28 日维持利率在 3.50%-3.75% 不变。 2026 年伊始，美股的叙事轨迹发生了微妙而坚决的转向。 一方面，创纪录的低失业金人数让美联储 1 月"按兵不动"的概率飙升至 95%；另一方面，台积 电的一席话和华尔街投行的亮眼财报，让市场从对 AI 泡沫的忧虑中惊醒。 美元走势：受此推动，美元指数走强。与之相对的是金价从 4630 美元高位回落，油价因地缘局势缓和与美元强势的双重挤压，跌幅超 4.8%。 这意味着，"放水驱动"的逻辑已成过去式，2026 年是实打实的"增长驱动"年。 牛市不再是虚无缥缈的预期，而是一场基于基本面的"多点开花"。 二、 叙事重心：AI 到底是不是泡沫？台积电给出答案 一、 宏观定调：强就业消解降息幻想 最新的数据给了市场一个"下马威"：截至 1 月 10 日当周，美国初请失业金人数降至 19.8 万，创下两年来第二低水平。 当华尔街还在纠结 AI 投资是否过热时，上游巨头台积电（TSM）用 560 亿美元的资本开支计划投下了信任票。 结论 ...

Core Insights - Nvidia's market capitalization has increased from approximately $345 billion in November 2022 to $4.5 trillion, making it the most valuable company globally [1] - The company is positioned for significant growth, with projections indicating it could reach a market cap between $7 trillion and $9 trillion by the end of 2026 [14] Group 1: Market Position and Growth Potential - Nvidia's backlog is reported to be around $500 billion, with the order backlog growing exponentially due to new deals, including a significant agreement with Anthropic [2][3] - Major partnerships, such as a $38 billion deal with OpenAI and a $20 billion licensing agreement with Groq, are expected to enhance Nvidia's revenue streams [4] - The company is projected to capture over 60% of big tech's infrastructure spending in AI, with potential data center revenue reaching between $320 billion and $330 billion by 2026 [6][7] Group 2: Valuation and Financial Metrics - Nvidia's current market cap is valued at approximately 27 times its trailing-12-month data center sales, which generated $167 billion over the past year [11][12] - If Nvidia doubles its data center sales while maintaining its valuation profile, it could be valued at nearly $9 trillion, or around $7 trillion if the market cap to sales ratio compresses [12] - The company's price-to-sales ratios have compressed over the past year, suggesting a shift in market perception towards a more mature business model [8][10]

Core Viewpoint - The automotive industry is shifting from reliance on NVIDIA's chips to self-developed solutions, as companies like Xiaopeng and NIO are moving towards their own chip development to reduce costs and enhance performance [2][5][20]. Group 1: Market Dynamics - Xiaopeng has released four new models equipped with its self-developed Turing driving chip, marking a complete departure from NVIDIA's chips [2]. - NIO is transitioning to its self-developed Shenqi chip, which is expected to significantly reduce costs compared to purchasing NVIDIA chips [2][8]. - NVIDIA's market share in China's high-end driving chip sector is projected to decline from 39% in 2024 to 25% in 2025, indicating a significant shift in the competitive landscape [2][9]. Group 2: Competitive Landscape - In 2024, NVIDIA's Orin-X chip held a 39.8% market share with 2.1 million units, but competitors like Horizon are gaining ground with lower-cost alternatives [5][6]. - Horizon's J5 chip has secured contracts with over nine automakers, including major brands like BYD and SAIC, highlighting the increasing competition in the market [6][9]. - By 2025, NVIDIA's partnerships with major clients like NIO and Xiaopeng have diminished, indicating a loss of influence in the market [6][8]. Group 3: Financial Performance - In the first three quarters of fiscal year 2025, NVIDIA reported $147.8 billion in total revenue, with only $1.7 billion (approximately 1%) coming from automotive business, underscoring the limited impact of automotive sales on overall performance [14][16]. - The automotive segment's revenue is significantly overshadowed by NVIDIA's data center business, which accounts for nearly 90% of its income [14][16]. Group 4: Future Outlook - NVIDIA is attempting to pivot from being a chip supplier to a comprehensive smart driving solution provider, but this transition faces challenges as competitors strengthen their positions [20]. - The launch of NVIDIA's open-source VLA model, aimed at assisting automakers lacking full-stack development capabilities, reflects its strategy to maintain relevance in the evolving market [20][21]. - However, the practical utility of the VLA model has been questioned, indicating potential hurdles in its adoption and effectiveness [21].