液冷及液冷工质市场更新 20251130 摘要 全球液冷市场快速增长，预计未来 3-5 年保持 20%-25%的年增长率， 2024-2025 年前三季度市场规模达 60-70 亿美元，北美占比最高，达 50%-55%。 北美头部数据中心通过新能源、储能和分布式供电等替代能源方案应对 电力瓶颈，但成本较高；国内厂商则通过采购前一代 GPU 芯片和东南亚 数据中心规避芯片限制。 高功率 GPU 系统的数据中心设计中，电系统采用 N+N 或 3+3+1 冗余 供电模式，热管理系统采用 N+1 冗余，冷板液冷的关键器件如循环水 泵也采用 N+1 冗余。 国内 AI 集群业务中，风冷与液冷共存，H100 液冷机柜通常采用 30% 风冷、70%液冷热板方式，单芯片功耗未超过 1,000 瓦，以单向液冷热 板为主。 冷板式和静默式液冷技术的选择依据 GPU 芯片热流密度，1 千瓦以内用 风冷，1-2 千瓦推荐单向液冷热板，超过 2 千瓦建议双向液冷热板，未 来 Ultra 系列或需转向双向相变方案。 Q&A 目前全球液冷市场的规模和区域分布情况如何？维谛技术在其中的市场份额是 多少？ 从 2025 年开始，随着英伟达 G ...

markets /macro 市场/宏观 a check-down of a dozen charts that stick out to me right now. 当前让我印象深刻的⼗⼏张图表⼀览 they don't carry a consistent theme nor bias, but I do think they speak to how wildly interesting the trading environment is these days. 这些图表虽未呈现统⼀主题或倾向，但确实展现了当下交易环境的极度趣味性 one thing struck me after assembling them: 整理完这些图表后，有⼀点令我颇为震撼： given the starting point of some scorching rallies in October ... and, for as high velocity as November was (see point #3 below) ... the fact that S&P finished this month in th ...

TPU 代工视角看谷歌材料 20251130 摘要 谷歌通过特殊 IP 和电源散热优化，在 2020 年帮助其数据中心 PUE 降 低约 25%，从而在 2020-2024 年成为其独家代工商，受益于谷歌链的 崛起。 谷歌与联发科联合设计自研芯片，博通专注于训练芯片，谷歌则进入推 理领域，GPT-8 亿芯片预计 2026 年 11 月上市，博通和联发科是主要 供应商，未来可能引入其他供应商。 自 2024 年 1 月起，伟创力加入谷歌制造业链条，与我们形成 80%对 20%的份额分配。2026 年起，红海将作为第三家供应商加入，分别占 据 65%、20%和 15%的市场份额，并提供专业液冷解决方案。 2025 年上半年，依数曾是谷歌最大 PCB 供应商，下半年切换回沪电， 目前沪电占 70%，方正占 20%，TTM 占 10%。光模块方面，旭创一直 是核心供应商，新易盛份额不到 10%。 谷歌计划从 2026 年开始，在交换部分切换到 AOC 加 LPO 组合，以降 低成本，这将导致原有光模块供应商结构性变化。线缆方面，从传统 AEC 线缆转向 AOC 线缆，国内长兴博创、海外菲尼特为主要供货方。 Q&A 谷 ...

导读 过去两年，小语言模型（SLM）在业界备受关注：参数更少、结构更轻，理应在真实部署中 "更快"。但只要真正把它们跑在 GPU 上，结论往往令人意外 —— 小模型其实没有想象中那么快。 参数缩小了，延迟却常常没有同步下降；结构轻量化了，吞吐却未必提升。这并非个别模型的问题，而是小模型设计长期忽略了一个根本事实： "模型更小" 并不 等于 "延迟更友好" 。 英伟达研究院就是从这一盲区重新出发：不是把大模型简单缩小，而是把 "真实 GPU 延迟" 作为结构设计的第一原则，全面重构小模型应该长成的样子。最终构建 的 Nemotron-Flash 模型同时实现了 SOTA 准确率、低延迟、高吞吐，打败了众多业界小模型。Nemotron-Flash 已集成进 TensorRT-LLM，单 H100 GPU 吞吐可达 41K tokens/second。 该论文已被 NeurIPS 2025 接收，相关海报将于 12 月 4 日在 San Diego 展示。 论文链接：https://arxiv.org/pdf/2511.18890 Hugging Face: https://huggingface.co/nvidi ...

本文来自格隆汇专栏：中信证券研究；作者：许英博 徐涛等 中美科技股大涨后回落，市场出现显著分歧： 1）AI Capex见顶了么？美股科技股见顶了么? 2）国产GPU/ASIC能否实现替代和突围？2026年中国AI算力会是投资机遇么？为了回答这些难题，中信证券研究部10余行业20余位分析师，持续贯穿中 美AI全产业链开展研究，面向2030年的全球和中国AI市场，推出《2030人工智能展望》系列报告。本报告是第一篇，我们创造性地构建了一套完整测算 框架：AI应用展望→token消耗量预测→需要的推理算力规模测算→需要的总算力规模→2030年新增算力规模→AI芯片市场空间展望。 我们认为，当前仍处于人工智能应用的早期。即便没有最终实现AGI，AI仍将释放互联网30年积累的数据红利，低成本提升知识和智能获取的便利性，从 而提高效率和生产力。我们预计，2030年的全球AI Capex有望在2025年的基础上增至5-7倍，中国AI Capex有望增至7-9倍。AI投资远没有见顶。但如美国 AI投资出现季度波动，仍可能导致高估值的美股科技股出现大幅波动。外资对中国科技股持股比例持续偏低，未来如中国AI投资超预期，或技术进步 ...

Core Viewpoint - The article emphasizes that artificial intelligence (AI) is transforming the global economy and presents a significant investment opportunity for investors through various fund options, particularly ETFs and public/private funds [1]. ETF Investment - ETFs are highlighted as an efficient tool for investors who prefer to follow industry trends without the hassle of selecting fund managers. The main focus of AI investment in A-shares is on "computing power infrastructure" and "application end" [2]. - Core broad-based ETFs include the AI ETF (515980) and AI ETF (515070), which track the China Securities Artificial Intelligence Index. These ETFs cover leading companies across the AI value chain, including chip manufacturers (e.g., Cambricon, Haiguang Information), large models and algorithms (e.g., iFlytek), and application scenarios (e.g., Hikvision, Kingsoft) [3][4]. - Segment-specific ETFs such as Cloud Computing 50 ETF (516630) and Communication ETF (515880) focus on computing power hardware and high-speed network facilities that support AI data transmission, respectively. The rationale is that hardware providers often see early performance returns in the AI development phase [5][6][7][8]. Public Funds (Active Equity) - Public funds rely on professional stock selection to seek alpha. The A-share market experiences rapid style rotation, and skilled fund managers can rotate investments within the AI value chain based on fundamental research [9][10]. - Focus on veteran managers in the "digital economy" and "TMT" sectors, particularly those with a track record during the mobile internet wave from 2013-2015. These managers tend to select companies with real performance rather than mere narratives [11][12]. - Quantitative public funds, such as those tracking the CSI 500 or CSI 1000 indices, excel in the active mid- and small-cap companies within the AI sector, often outperforming benchmark indices [13]. Private Funds - Private funds are characterized by greater flexibility in position management and the use of derivatives for risk hedging. They can effectively manage volatility in the AI sector by controlling drawdowns during declines and capitalizing on gains during upswings [14][15]. - Notable institutions include Huanfang Quantitative, Jiukun Investment, and Yifan Investment, which leverage deep learning to uncover market patterns and opportunities that active management may overlook [16]. - The article also highlights the importance of investing in global AI leaders through local private funds, as the U.S. maintains a dominant position in high-end computing and foundational models [18]. Recommended Fund Analysis - The Keywise Penguin No. 1 fund is recommended for its strong reputation and global investment scope, covering major tech markets and key AI players like Nvidia, Microsoft, and TSMC. The fund's strategy includes both long and short positions to protect net value during market fluctuations [19][20][21]. Investment Strategy Summary - The article concludes with a tailored investment strategy for different investor types, recommending ETFs for conservative investors, public funds for those seeking alpha, and the Keywise Penguin No. 1 for high-net-worth individuals looking for global exposure to AI assets [22].

>>自动驾驶前沿信息获取 → 自动驾驶之心知识星球 论文作者 | Haohong Lin等 编辑 | 自动驾驶之心 英伟达最近工作很多啊，而且做的都挺扎实。 前一段时间的自驾VLA框架 - Alpamayo-R1，昨天新的一篇闭环仿真测试框架 - MPA。 可圈可点，今天自动驾驶之心 为大家分享的就是这篇新工作MPA。 自动驾驶中的开环评测已经相对完善，但在闭环评测中仍然面临着级联误差和泛化能力不足的问题。针对这个问题，CMU、斯坦福和英伟达的团队提出一种基于模型 的策略自适应通用框架 - Model-based Policy Adaptation。旨在提升预训练E2E驾驶智能体在部署阶段的鲁棒性与安全性。MPA首先利用几何一致的仿真引擎生成多样化 反事实轨迹，让智能体接触到原始数据集之外的场景；基于生成的数据，MPA训练一个基于扩散模型的策略适配器以优化基础策略的预测结果，并训练一个多步Q值 模型来评估长期收益。推理阶段，适配器生成多个轨迹候选，Q值模型则选择期望效用最高的轨迹。在nuScenes基准数据集上，通过重建出的真实闭环仿真器的实验 表明，MPA在域内场景、域外场景及安全关键场景中均显著提升了性能 ...

如果说三年前ChatGPT的横空出世，是推倒多米诺骨牌的第一指，那么三年后的今天，我们正身处这场骨牌相继倒地的连锁震荡中。在这一波 几乎每天都在更新迭代的技术浪潮中，大多数企业管理者正感到前所未有的迷茫：风口变幻莫测，战略究竟该怎么定？ 在ChatGPT诞生三周年的今天，中欧国际工商学院战略学教授张宇，并未随波逐流地预测技术走向，而是从商学院教授的视角，回归商业本 质，梳理出AI时代战略制定中那些"不变"的底层逻辑：在这个充满不确定的时代，不妨先确立那些确定的事。 毫无疑问，过去三年中，以生成式AI（Generative AI）为代表的新一代AI技术，给人类社会带来了深远改变，从大幅提升工作速度（尤其是入门级的图文 和编程工作），到节约工作时间，再到节约成本和提升利润，以及威胁和替代工作岗位（尤其是初级岗位和年轻人的就业机会），让我们几乎每天都感到 震惊与惊叹。 时至今日，恐怕绝大多数人都会同意，这一波AI技术是自蒸汽机、电力、计算机/互联网之后，对人类社会产生深远影响的"第四次技术革命"。 那么，为什么生成式AI会对我们的社会尤其是商业活动产生如此巨大的影响和改变呢？究其原因，恐怕是因为生成式AI是迄今为止 ...

Core Insights - Berkshire Hathaway currently holds stakes in two quantum computing stocks: Alphabet and Amazon, reflecting Warren Buffett's long-term investment strategy that has yielded a compound annual gain of 20% over 60 years, nearly double that of the S&P 500 [1][2] Investment Philosophy - Warren Buffett is known for his contrarian investment approach, avoiding hype-driven stocks that often lead to overstretched valuations [3] - Buffett's portfolio includes major positions in companies like Apple, American Express, Bank of America, Coca-Cola, and Chevron, showcasing a diversified strategy across various sectors [4] - The companies in Buffett's portfolio are resilient, generating consistent cash flow that is reinvested or returned to shareholders [5] Technology Sector Focus - Technology stocks represent a smaller portion of Berkshire's portfolio due to their higher valuation multiples and rapid changes in the industry [6] - Berkshire has invested in technology, particularly in artificial intelligence (AI), with significant positions in Amazon and Alphabet [7] Company Analysis - Apple has established a strong customer lock-in through its hardware and services, while Amazon has become a leading online marketplace and diversified into various sectors [8] - Alphabet has leveraged its expertise in internet search to develop new services relevant to its AI initiatives, generating steady cash flow for innovation and shareholder rewards [9] Potential Investment in Nvidia - There is speculation that Berkshire may acquire a stake in Nvidia, which aligns with Buffett's investment criteria due to its strong brand in semiconductors and AI infrastructure, as well as its modest dividend and stock buyback strategy [10][12] - Nvidia's ecosystem supports generative AI development and plays a crucial role in hybrid classical-quantum computing environments, positioning it for growth as the AI narrative evolves [13] - Currently, Nvidia trades at a forward price-to-earnings (P/E) multiple of 24, which is considered a premium but is the lowest price in over a year, with accelerating revenue and profits [14][15]

近期，手术机器人行业掀起"英伟达合作热潮"，目前已有超8家手术机器人企业宣布牵手英伟达。 例如，内腔机器人企业EndoQuest Robotics计划将英伟达的IGX Thor平台集成到其下一代手术机器人系统中；英 国CMR Surgical宣布成为英伟达IGX Thor平台的全球首批应用企业。据了解，IGX Thor是英伟达最新推出的顶级 物理AI与机器人平台，具备5581 FP4 TFLOPS人工智能算力与400GbE网络连接能力，较前代平台实现了算力能 效与感知能力的全面飞跃。 此外，强生、XCath、Asensus、Moon Surgical、Virtual Incision、Neptune Surgical、Stereotaxis等手术机器人企业 均与英伟达达成了合作。 | 手术机器人企业 | 与英伟达的合作内容 | | --- | --- | | EndoQuest Robotics | 将英伟达的IGX Thor平台集成到其下一代手术机器人系统 | | | 日 | | CMR Surgical | 成为英伟达IGX Thor平台的全球首批应用企业 | | 强生 | 将英伟达的Isaac for ...