策略研究 | 证券研究报告 — 点评报告 2025 年 7 月 13 日 策略点评 国产算力产业链贯通，行情催化在即 国产算力产业链"产业突破-业绩验证-需求验证"逻辑链条已逐步打通，产业 链行情正在迎来催化。 tianran.gao@bocichina.com 证券投资咨询业务证书编号：S1300522100001 ◼ 国产算力产业链"产业突破-业绩验证-需求验证"逻辑链条已逐步打通，产 业端，沐曦集成、摩尔线程两大国产 GPU 厂商科创板 IPO 获受理，填补 A 股全功能 GPU 空白，华为算力集群性能媲美英伟达，产业链重点公司 业绩来看，工业富联中报业绩创历史同期新高，需求端，国产大模型 Tokens 消耗量高速增长。国产 GPU 上市潮、硬件厂商业绩兑现与 AI 大 模型需求增长形成共振，国产算力产业从技术研发、商业落地到需求支撑 已进入高速增长周期，全产业链自主化进程提速，国产算力行情亦有望受 到催化，建议关注国产算力芯片、服务器、PCB、光通信厂商等。 ◼ 风险提示：1）政策落地不及预期，宏观经济波动超预期。2）市场波动风 险。3）海外经济超预期衰退、流动性风险。 策略点评 国产算力产业链"产业 ...

Core Viewpoint - The AI investment cycle is transitioning from "investment" to "returns," but this does not imply a slowdown is a peak. Goldman Sachs indicates that despite a deceleration in growth, AI infrastructure investment will remain sustainable over the next 2-3 years, with cost benefits already being realized and stock prices not yet reflecting this structural change [1][2]. Group 1: AI Investment and Returns - Goldman Sachs categorizes AI value creation into three phases: cost reduction through automation (current phase), reinvestment and rebuilding, and revenue generation through incremental income [2][3]. - AI applications in customer service, sales, and IT are already yielding tangible benefits, with 43% of call centers adopting AI tools and achieving an average operational cost reduction of 30% [2][3]. Group 2: Cost Savings and Future Projections - By 2030, AI automation could save Fortune 500 companies approximately $935 billion, representing about 14% of their total costs, with a net present value return of around $780 billion against a cumulative investment of $350 billion [3][4]. - Major cloud service providers are the primary investors in AI infrastructure, focusing on long-term revenue growth opportunities rather than short-term cost savings, complicating ROI calculations [3][4]. Group 3: Infrastructure Spending and Demand - Concerns about whether infrastructure spending has peaked, particularly regarding training chip inventory and demand, are considered overstated by Goldman Sachs [4][5]. - Large tech companies like Microsoft, Amazon, Google, and Meta are expected to maintain their AI infrastructure investments without significantly compressing profit margins over the next 2-3 years [5][6]. Group 4: New Demand Drivers - Demand for "inference" computing from enterprise clients and government (sovereign AI) is emerging as a new spending driver, especially as small and medium enterprises rapidly expand their deployment of customized models or edge AI applications [6][7]. Group 5: Market Valuation and Stock Performance - The market has partially priced in strong demand expectations for Nvidia's next-generation GPUs, but there is still insufficient valuation for its expanding customer base and the potential explosion of AI inference business [8]. - Broadcom's stock price increase is attributed to clear guidance indicating AI revenue growth of 60% in FY25 and FY26, suggesting that the stock price rise reflects a clearer mid-term fundamental improvement path [8].

July 8, 2025 12:27 PM GMT Greater China Technology Hardware | Asia Pacific GB200 NVL72 racks in June We estimate that GB200 rack output increased slightly in June, to ~2.5k vs. 2-2.5k in May, leading total 2Q GB200 rack output to reach ~6k. We expect shipments to continue to increase in 3Q; they appear to be on track to reach 10-11k. Key Takeaways Quanta (2382.TW, OW): June revenue was ~NT$190bn (+18% m/m, +71% y/y). We attribute the m/m increase to: We estimate that Quanta's GB200 racks increased from ~400 ...

Group 1 - Amazon's cloud computing division has developed specialized hardware for cooling next-generation NVIDIA GPUs, which are widely used for AI-related computing tasks [1] - The high energy consumption of NVIDIA GPUs necessitates additional cooling equipment for companies utilizing these processors [1] - Amazon previously considered building data centers with liquid cooling systems but found existing solutions inadequate for their scale [1] Group 2 - The newly developed In-Row Heat Exchanger (IRHX) can be integrated into existing and new data centers to support high-power NVIDIA GPUs [1] - Customers can now access a new AWS service through the P6e compute instance, which utilizes NVIDIA's high-density computing hardware [2] - Amazon has reduced reliance on third-party suppliers by developing its own infrastructure hardware, contributing to improved profitability [2]

行 业 及 产 业 计算机 2025 年 07 月 09 日 相关研究 《从 CloudMatrix 看超节点趋势！3D 打 印产业强趋势！——计算机行业周报 20250616-20250620》 2025/06/21 《字节 FORCE 大会亮点！华为 HDC 2025 关注点！——计算机行业周报 20250609-20250613》 2025/06/14 证券分析师 黄忠煌 A0230519110001 huangzh@swsresearch.com 李国盛 A0230521080003 ligs@swsresearch.com 林起贤 A0230519060002 linqx@swsresearch.com 杨海晏 A0230518070003 yanghy@swsresearch.com 洪依真 A0230519060003 hongyz@swsresearch.com 刘洋 A0230513050006 liuyang2@swsresearch.com 研究支持 曹峥 A0230525040002 caozheng@swsresearch.com 崔航 A0230524080005 cuihang@ ...

This achievement continues CoreWeave's legacy of delivering first-to-market access to the world's most advanced AI infrastructure demanded by the world's leading AI labs and enterprises. This initial deployment of NVIDIA GB300s expands on CoreWeave's existing Blackwell fleet, which also includes the NVIDIA HGX B200 and the NVIDIA GB200 NVL72 system. Last year, CoreWeave was among the first to offer NVIDIA H200 GPUs and was the first AI cloud provider to make NVIDIA GB200 NVL72 systems generally available. I ...

IDC 在其最新的全球季度服务器追踪报告中估计，基于 Arm 架构的服务器今年将占全球总出货量的 21.1%，而不是Arm 基础设施主管 Mohamed Awad在 4 月份宣称的 50%。 公众号记得加星标⭐️，第一时间看推送不会错过。 来源：内容 编译自 theregister 。 基于 Arm 的服务器在市场上迅速获得关注，预计 2025 年出货量将增长 70%，然而，这仍然远远低 于该芯片设计商的目标，即到今年年底占据全球数据中心 CPU 销售额的一半。 市场观察机构 IDC 表示，Arm 服务器之所以受到广泛关注，主要得益于大型机架规模配置的推出， 例如专为AI 处理而设计的Nvidia DGX GB200 NVL72等系统。 配备至少一个GPU的服务器（有时被称为支持AI）预计将增长46.7%，几乎占今年市场总价值的一 半。超大规模客户和云服务提供商的快速采用正在推动服务器市场的发展，IDC表示，服务器市场规 模将在短短三年内增长两倍。 2025年第一季度，服务器市场整体规模达到创纪录的952亿美元，同比增长134.1%。因此，IDC将全 年服务器市场规模预测上调至3660亿美元，增幅达44.6% ...

Core Insights - CoreWeave, Inc. (CRWV) is launching advanced offerings tailored for AI model training and inference, positioning itself as a competitive player in the Infrastructure-as-a-Service market with a projected global economic impact of $20 trillion from AI by 2030 and a total addressable market (TAM) of $400 billion by 2028 [1] Product Development - The company has introduced three new AI cloud software products aimed at accelerating AI development processes, marking its first software release since acquiring Weights & Biases in May 2025 [2][8] - New products include Mission Control Integration for rapid training issue resolution, W&B Inference for testing open-source AI models, and Weave Online Evaluations for real-time performance feedback [2] Technological Advancements - CRWV has rapidly adopted cutting-edge technology, being the first to deploy NVIDIA's H100 and H200 GPUs at scale and to offer GB200 NVL72 instances, which has positively impacted its revenue [3] - The introduction of next-gen AI Object Storage, combined with Kubernetes services, provides a comprehensive production-level setup for AI customers from the outset [3] Strategic Focus - The company is concentrating on four main areas: scaling capacity, financing infrastructure, enhancing platform differentiation, and expanding market reach [4] - CoreWeave is expanding its global presence to tap into new markets and strengthen existing customer relationships [4] Competitive Landscape - Microsoft (MSFT) is a major competitor in AI infrastructure, leveraging its Azure platform and significant investments in OpenAI to enhance its market position [5] - Nebius Group N.V. (NBIS) is also emerging as a specialized AI infrastructure provider, focusing on large-scale GPU clusters and cloud platforms, supported by its partnership with NVIDIA [6] Financial Performance - CoreWeave's stock has surged by 334.2% year-to-date, significantly outperforming the Internet Software industry's growth of 13.1% [7] - The company trades at a forward price-to-sales ratio of 10.31X, which is higher than the industry average of 5.68 [9]

Nebius (NASDAQ: NBIS ) has accomplished in eight weeks what larger cloud companies required far longer to attempt. The company combined a US$1 billion two-tranche convertible note priced on June 5 with immediate deliveries of NVIDIA GB200 NVL72 and"AWS Certified AI Practitioner Early Adopter"I am a DevOps Engineer for a major, wholly owned subsidiary of a large-cap Fortune 500. I am a true subject-matter expert on the actual buildout, deployment, and maintenance of AI tools and applications. I have increasi ...

每经记者｜王晶 每经编辑｜马子卿 "昇腾算力到底能不能给大家底气和信心，有些人将信将疑，甚至认为昇腾训练不出先进的大模型。"华 为专家开场的这句话，揭示出人们对国产AI（人工智能）算力的焦虑：在美国芯片禁令的阴影下，这 支国产算力舰队究竟能否突破算力封锁、重构AI产业格局？ 答案正随着盘古Pro MoE模型、盘古Ultra MoE模型、昇腾384超节点等各种实践的落地而变得清晰。 近日，记者从一位华为专家处获悉，华为发布的昇腾384超节点已经开始发货，这不仅是目前业界规模 最大的超节点，更是一项技术突破：华为将384颗昇腾AI芯片连接在一起组成了集群，通过全对等高速 互联的架构，基于系统工程的方法，对计算、内存、通信的极致优化调度，可提供高达300 PFLOPs的 密集BF16算力，该性能表现接近英伟达GB200 NVL72系统的两倍。 这引出一个关键问题：在单芯片工艺暂时落后的情况下，昇腾是如何实现算力超越的？ "集群"绝非简单的"芯片堆叠"，而是华为在极端压力下，以系统工程思维进行的一次"突围"。"华为内 部有个算力会战，把华为云、模型、底座、芯片、硬件工程、基础软件的人集结在一起，深度协同。打 造复杂、 ...