想要看外资研报的同学，给大家推荐一个星球，在星球中每天都会上传几百篇外资顶尖投行的原文研 报：大摩、小摩、UBS、高盛、Jefferies、HSBC、花旗、BARCLAYS 等。 还有专注于半导体行业分析的SemiAnalysis的全部分析报告： 星球中每日还会更新Seeking Alpha、Substack、 stratechery的精选付费文章， 现在星球中领券后只需要 390元，即可每天都能看到上百篇外资顶尖投行科技行业的分析报告和每天的精选报告，无论是我们自 己做投资，还是对行业有更深入的研究，都是非常值得的。 ...

路透社5月9日报道，三位知情人士透露，在美国对原版芯片实施出口限制后，英伟达计划在未来两个月 内为中国市场推出其 H20 人工智能芯片的降级版本。 其中两位消息人士称，这家美国芯片制造商已通知中国主要客户（包括领先的云计算供应商）， 其目 标是在 7 月推出改良后的 H20 芯片 。 这款降级的 H20 芯片代表了英伟达在面对美国政府扩大限制中国获取先进半导体技术的努力下，维持 其在关键市场地位的最新尝试。 H20 原本是英伟达获准在中国销售的最强人工智能芯片，但美国官员上月通知该公司称，该产品需要出 口许可证，实际上已被禁止进入市场。 在截至 1 月 26 日的财年中，中国市场为英伟达贡献了 170 亿美元收入，占其总销售额的 13%。 凸显中国战略重要性的是，英伟达首席执行官黄仁勋上月访问了北京—— 就在美国官员宣布对 H20 芯 片实施新出口许可要求的几天后。 黄仁勋在与中国官员的会面中强调，中国是公司的关键市场。 自 2022 年以来，美国以担忧芯片可能用于军事用途为由，限制向中国出口英伟达最尖端的芯片。H20 芯片是在 2023 年 10 月美国收紧出口管制后推出的。 据路透社今年早些时候报道，随着 ...

最近民主党众议员Bill Foster（伊利诺伊州）计划在未来数周提交法案，要求受出口管制的美国先进AI 芯片必须配备定位追踪和远程禁用功能。根据路透社报道，这位粒子物理学家出身的议员旨在通过技术 手段（如英伟达等芯片制造商可实时追踪芯片位置，并在芯片进入未授权区域时远程禁用）遏制对华非 法出口。法案矛头直指中国，媒体此前已多次报道英伟达AI芯片通过灰色渠道流入中国市场。 我们认为即使法案最终立法，对中国AI资本支出和IDC需求的乐观预期影响有限。近期中国AI产业调研 进一步验证：随着DeepSeek等大模型的突破，中国AI商业化进程（涵盖ToB和ToC应用）已进入快车 道。未来数年，本土7纳米晶圆厂产能扩张将成为算力基建的核心驱动力。预计2026年起随着B/C端应 用爆发，推理需求将主导增长，而该领域对大规模GPU集群和CUDA生态依赖度较低——本土芯片甚至 游戏GPU即可满足需求。 技术可行性存疑 从技术层面看，定位追踪功能较易实现，英伟达或已具备相关技术储备。但远程禁用功能的有效性存在 争议。即便技术可行，此类系统也存在被黑客攻破的风险，导致管控措施形同虚设。因此，该法案虽会 增加美国AI芯片企业的合规 ...

Core Viewpoint - The article discusses the performance of domestic GPUs, highlighting the competitive landscape and the impact of sanctions on production capabilities. It emphasizes that while the BR100 from Birun is the most powerful, it cannot be mass-produced due to sanctions, leaving Huawei's 910C as the strongest available option [1][2]. Performance Comparison - The performance ranking of GPUs indicates that the Huawei 910C leads, followed by the Haiguang BW100, with the latter achieving approximately 87% of its potential performance. The Cambrian 590's performance is uncertain, potentially reaching 80% [2]. - The Haiguang BW100 has a deep computing capability of over 400T, which is about 50% of the H800's performance, while the 910C reaches approximately 60% of the H800's performance [1][2]. Memory and Bandwidth - Most domestic GPUs utilize HBM2e memory due to sanctions preventing the use of HBM3e. Sufficient memory is crucial for initiating inference tasks, as insufficient memory can hinder model startup [3]. - Huawei leads significantly in memory bandwidth, which is a critical factor for performance [4]. Pricing Insights - Current market prices for GPUs are as follows: Haiguang BW100 at approximately 100,000 yuan, Huawei 910B at around 70,000 yuan, and 910C at about 180,000 yuan. The Cambrian 590 has seen a price drop from 80,000-85,000 yuan to between 60,000-70,000 yuan [2]. Technical Specifications - A table summarizes key specifications of various GPUs, including memory type, bandwidth, capacity, and arithmetic intensity. For instance, the Huawei 910C features HBM2e with a bandwidth of 3.2 TB/s and a memory capacity of 64 GB [5].

Core Viewpoint - The "2025 Second High-Performance Chip Developer Forum and Chip Thermal Management Technology Exchange Conference" will be held on May 22-23 in Beijing, focusing on various aspects of domestic AI chip development and technology [1] Group 1: Conference Details - The forum will cover topics such as domestic AI chip progress, AI chip security, Chiplet technology, advanced packaging materials, thermal design for AI chips, direct cooling technology, and 3DVC uniform temperature technology [1] - Over 20 presentations are scheduled, with an expected attendance of more than 300 industry experts [1] Group 2: Participation and Services - Attendees will receive learning opportunities, including post-conference materials and access to a dedicated community [1] - Exhibition services include booth display, three attendee passes, printed conference materials, public account promotion, and video interviews [1] - Speaking opportunities include a 30-minute presentation in a specialized session, three attendee passes, printed conference materials, and public account promotion [1]

星球中每日还会更新Seeking Alpha、Substack、 stratechery的精选付费文章， 现在星球中领券后只需要 390元，即可每天都能看到上百篇外资顶尖投行科技行业的分析报告和每天的精选报告，无论是我们自 己做投资，还是对行业有更深入的研究，都是非常值得的。 还有专注于半导体行业分析的SemiAnalysis的全部分析报告： 想要看外资研报的同学，给大家推荐一个星球，在星球中每天都会上传几百篇外资顶尖投行的原文研 报：大摩、小摩、UBS、高盛、Jefferies、HSBC、花旗、BARCLAYS 等。 ...

总体而言，中国半导体自给率为 24%，高于我们的预期。根据高德纳（Gartner）的数据，中国半导 体公司的总收入为 430 亿美元，较 2023 年的 320 亿美元增长了 36%。中国半导体市场规模约为 1830 亿美元，占全球需求的 29%。因此，中国半导体自给率约为 24%，比 2023 年的 20% 提高了 4 个百分点。我们认为，中国半导体自给率的提升主要归因于以下几点：第一，政府对家电和消费电 子产品的补贴刺激了需求，与全球同行相比，中国半导体供应商在消费市场的曝光度相对较高；第 二，大多数中国半导体公司在 2023 年底或 2024 年上半年完成了库存消化，因此 2024 年中国半导体 公司的增长超过了外国同行；第三，存储芯片和先进节点产品（智能手机系统级芯片、GPU 等）的 产能提升和良率改善。 这篇文章是Morgan Stanley对中国GPU自给自足程度的研报的部分内容，最后那张图很有意思，把华 为昇腾系列GPU在2024到2027（预测）的产能、良率、价格、产量和收入情况都详细列了出来，相 信很多朋友对这个数据也都很有兴趣。报告原文放星球了。 正文 据我们估算，2024 年中国人工智能 ...

这是substack上的一篇分析，作者是Mary Clare McMahon，发布于ChinaTalk，我们之前发过几篇 ChinaTalk的文章，写的都很好，而且很多都是免费的，推荐大家看一下。本篇文章已获得作者的转 载许可。原文链接如下： 英伟达的软件护城河 1. 构建自己的软件堆栈，包括专有的并行编程模型和开发者依赖的周边工具，以高效编写、优化和 部署代码。 2. 深化与PyTorch的整合，PyTorch是最广泛采用的开源机器学习框架，用于模型训练。 3. 投入工程资源开发开放神经网络交换（ONNX），这是一个支持跨硬件平台模型可移植性的开放 标准，以支持在昇腾芯片上部署非昇腾训练的模型。 https://www.chinatalk.media/p/can-huawei-compete-with-cuda 这篇文章的作者并不是中国人，但看了这篇分析会发现，虽然华为目前在挑战英伟达上面临着各种 各样的问题，但作者同样对华为充满了信心。 正文 上个月，有报道称DeepSeek在其R1推理模型的精简版上使用华为昇腾芯片运行。虽然DeepSeek的模 型是在英伟达Hopper系列芯片上训练的，但华为部署Dee ...

Group 1 - The article recommends a platform where users can access hundreds of foreign investment bank research reports daily, including those from top firms like Morgan Stanley, JPMorgan, UBS, Goldman Sachs, Jefferies, HSBC, Citigroup, and Barclays [1] - The platform also provides comprehensive analysis reports focused on the semiconductor industry from SemiAnalysis, along with selected paid articles from Seeking Alpha, Substack, and Stratechery [3] - The subscription to the platform is currently available for 390 yuan, offering access to a wealth of technology industry analysis reports and selected articles daily, which is deemed valuable for both personal investment and deeper industry research [3]

Market Overview - The global AI server market is projected to reach $125.1 billion in 2024, $158.7 billion in 2025, and $222.7 billion by 2028, with generative AI servers expected to increase their market share from 29.6% in 2025 to 37.7% in 2028 [1] Major Players' Shipment Data - NVIDIA holds nearly 70% of the AI chip market share, with the new Blackwell platform expected to account for 82% of its high-end GPU shipments in 2025. The company plans to launch the B30 for the Chinese market in the second half of 2025, with the B300 and GB300 expected to contribute 60-65% of its total GPU shipments for the year [2][3] - AMD's high-end GPU shipments are forecasted to grow by 48% in 2025, reaching approximately 585,000 units. The MI325 series has a low adoption rate, while the MI350 series is expected to enter the market in the second half of 2025 [5][4] - Intel's Gaudi3 high-end AI chip is projected to have a shipment volume of around 100,000 units in 2025, targeting CSPs and IBM as primary customers [6] - Google is expected to lead in cloud service provider (CSP) shipments with approximately 2.2 million TPUs in 2025, while AWS's self-developed ASIC shipments are projected to reach over 1.8 million units, nearly doubling from previous figures [7][8] Chip Type and Shipment Forecast - NVIDIA's high-end GPU shipments are expected to total around 6.6 million units in 2025, with significant contributions from the Blackwell platform [3] - AMD's MI series is expected to see a total shipment of 585,000 high-end GPUs in 2025, with the MI350 series anticipated to compete directly with NVIDIA's offerings [5] - The Ascend (昇腾) ASIC is projected to reach 450,000 units in 2025, driven by domestic AI demand and trade restrictions [8]