Vi e w p o i n t | 25 Feb 2026 19:27:39 ET │ 10 pages Taiwan Semiconductors Nvidia result implications - Blackwell expanding strongly, entering the Rubin Cycle CITI'S TAKE Nvidia just published its Jan Q result (report) and provided a positive outlook for the upcoming Rubin product cycle. While some investors were skeptical on Rubin's schedule and initial yield rate, Nvidia reiterates that the chip already has mass volume production and system rack is to ramp up big volumes in 2H26. Yet we believe the major ...

英伟达最新公布的季度业绩与未来财测全面击溃华尔街预期，高盛在最新研报中明确指出，这家芯片巨 头在未来数月内跑赢大盘的路径已变得异常清晰。 受超大规模云服务商持续强劲的资本支出推动，英伟达第一季度的营收指引大幅超出市场共识。据追风 交易台消息，高盛分析师 James Schneider 及其团队重申对该股的"买入"评级，并维持250美元的目标 价，这意味着该股较当前水平仍有近28%的上涨空间，此举预计将进一步提振市场对整个人工智能基础 设施板块的投资信心。 市场乐观情绪的背后不仅仅是历史业绩的兑现。高盛在报告中前瞻性地指出了推动英伟达持续走强的三 大核心催化剂：超大规模企业资本支出的上调预期、AI初创企业融资完成后的支出能见度跃升，以及 基于新一代架构的AI模型发布将再次印证其技术护城河。 此外，英伟达近期与 Meta、OpenAI 和 Anthropic 等顶尖科技巨头达成的深度战略合作与百亿美元级投 资布局，不仅从根本上锁定了未来的订单基本盘，也为包括存储和半导体设备在内的全球科技供应链带 来了广泛的积极溢出效应。 业绩与指引双双击溃市场预期 英伟达第四季度实现营收681亿美元，不仅高于高盛预期的673亿美 ...

Core Insights - Neurophos, an AI chip startup backed by the Gates Frontier Fund, has achieved a significant breakthrough in silicon photonics with its optical processing unit (OPU), which is approximately 10,000 times smaller than existing technologies and features a 1000×1000 pixel scale photonic computing matrix on a single chip [1][3] Group 1: Technology Advancements - The first optical accelerator, Tulkas T100, boasts AI computing performance that is ten times that of NVIDIA's latest Vera Rubin NVL72 supercomputer at FP4/INT4 precision, while maintaining similar power consumption levels [3] - Key technological advancements include a 1000×1000 photonic tile, significantly larger than the current GPU standard of 256×256, and a clock frequency of 56 GHz, which is much higher than the 9.1 GHz of Intel's Core i9-14900KF and 2.6 GHz of NVIDIA's RTX Pro 6000 GPU [3] Group 2: Production and Challenges - The CEO of Neurophos stated that traditional silicon photonic transistors are about 2 millimeters long, making high-density integration difficult, but their technology miniaturizes these transistors to a scale compatible with CMOS processes, enabling large-scale parallel optical computing [3] - Despite the advancements, the technology is still in the engineering validation stage, with mass production not expected before 2028, and several challenges remain, including on-chip SRAM capacity, vector processing unit expansion, and optoelectronic co-design [4]

Company Performance - Nebius Group NV (NASDAQ:NBIS) experienced a significant decline, falling 8.68% to close at $99.29, reflecting overall market pessimism due to geopolitical tensions triggered by President Trump [1][3] - The decline in Nebius's stock price was part of a broader market trend, with the tech-heavy Nasdaq index dropping by 2.39% [3] Geopolitical Impact - President Trump announced plans to impose 10% tariffs on goods from eight countries opposing his takeover of Greenland, prompting the EU to threaten retaliation with $108 billion in levies [2] Technological Developments - Nebius Group NV plans to deploy the NVIDIA Rubin platform through Nebius AI Cloud and Nebius Token Factory, aiming to enhance AI capabilities for its customers starting in the second half of the year [4] - As an NVIDIA Cloud Partner, Nebius will be among the first AI cloud providers to offer this platform across its full-stack infrastructure in the US and Europe [4] Market Outlook - Despite the current challenges, there is potential for Nebius as an investment, although some analysts believe other AI stocks may offer better returns with lower risk [6]

Investment Rating - The industry investment rating is "Overweight" (maintained) [1] Core Insights - The report highlights the expansion of storage demand driven by NVIDIA's new storage architecture, with continuous price increases in the semiconductor upstream [3] - The introduction of TSMC's 2nm mass production accelerates global resonance, opening up opportunities in fabs and equipment [3] - The AI terminal is initiating a new cycle of industrial chain upgrades in consumer electronics and components [3] Summary by Sections Industry Overview - The electronic industry is projected to outperform the overall market [1] - The report includes various related research reports that provide insights into the semiconductor and AI sectors [3] Technological Innovations - NVIDIA's Rubin platform showcases a collaborative design that significantly reduces training time and inference costs, marking a revolutionary change in architecture [4] - The Rubin NVL72 cabinet-level solution integrates 18 computing trays, achieving a total of 72 GPUs and 36 CPUs, with a CPU-GPU interconnect bandwidth of 1.8TB/s [5] - The system features high bandwidth, cable-free architecture, and deep interconnect capabilities, enhancing AI computing infrastructure [6] Beneficiary Companies - Beneficiary companies identified include: - Assembly: Industrial Fulian - Connectors: Luxshare Precision, Huizhu Technology, and Ruikeda - PCBs: Shenghong Technology, Huitian Technology, and Jingsheng Electronics - Cooling: Invec, Lingyi Zhi Zao, and others - Power Supply: Oulu Tong and Aohai Technology [7]

Core Insights - NVIDIA is fully committed to AI, marking its first appearance at CES in five years without showcasing gaming graphics cards [2] - The next-generation Rubin architecture GPU demonstrates significant performance improvements, with inference and training capabilities being 5 times and 3.5 times that of the Blackwell GB200, respectively [4][17] Group 1: New Product Launches - NVIDIA introduced five new product lines, emphasizing the importance of open-source training frameworks and multimodal datasets, including 100 trillion language training tokens and 100TB of vehicle sensor data [5][6] - The Vera Rubin NVL72 architecture was officially launched, featuring six core components designed to enhance AI data center capabilities [14][15] - The Rubin GPU achieves 50 PFLOPS in inference performance and 35 PFLOPS in training performance under NVFP4 data types, significantly surpassing previous models [17] Group 2: Technological Advancements - The NVLink 6 technology enhances inter-GPU bandwidth to 3.6 TB/s, with a total bandwidth of 260 TB/s across the entire architecture [21][20] - The Vera CPU integrates 88 custom Arm cores, allowing for high thread concurrency and improved memory bandwidth [22] - NVIDIA's new BlueField-4 DPU introduces a memory layer aimed at optimizing key-value cache operations, addressing performance bottlenecks in AI infrastructure [32][34] Group 3: AI Model Developments - The Alpamayo model series was launched for autonomous driving, featuring a 10 billion parameter open-source model capable of interpreting environmental data for decision-making [39][41] - The Nemotron model family expands into voice, retrieval-augmented generation (RAG), and safety applications, enhancing AI capabilities in various domains [49][51] - The Cosmos platform for robotics has been upgraded, providing new models for generating synthetic data that adheres to physical laws [54][58] Group 4: Healthcare and Life Sciences - NVIDIA Clara targets the healthcare sector, aiming to reduce costs and accelerate the implementation of treatment solutions [62] - The company offers a dataset of 455,000 synthetic protein structures to support research in drug discovery and personalized medicine [66][69]

Core Insights - Nebius will start offering NVIDIA's Vera Rubin NVL72 GPU in the second half of 2026 in the US and Europe, becoming one of the first AI cloud providers to deploy this computing platform [1] - The integration of Vera Rubin NVL72 into Nebius's full-stack infrastructure will enable customers to build next-generation AI applications with regional availability and control [1] - Nebius Token Factory is described as a post-training platform for enterprises, aimed at facilitating the development of AI systems [1] Product and Technology Overview - The Rubin computing platform will complement Nebius's existing NVIDIA GB200 NVL72 and Grace Blackwell Ultra NVL72 capacities, expanding customer options [2] - NVIDIA's CEO Jensen Huang announced that Vera Rubin has entered full-scale production, positioning it as the successor to Grace Blackwell [2] - Vera Rubin is characterized as an AI supercomputer composed of six core components, including Vera CPU, Rubin GPU, and others, designed for next-generation AI workloads in cloud and large data centers [2] Performance and Cost Efficiency - The Rubin GPU features a third-generation Transformer engine with NVFP4 inference performance of 50 PFLOPS, which is five times that of the previous Blackwell GPU [3] - The Vera Rubin platform can train large-scale Mixture of Experts (MOE) models in the same training time while requiring only a quarter of the GPUs, reducing the training cost per token to one-seventh of the previous generation [3] - Vera Rubin will support third-generation confidential computing technology, becoming the industry's first rack-level trusted computing platform, catering to AI scenarios with high demands for security isolation, data privacy, and multi-tenant environments [3]

Core Viewpoint - NVIDIA is shifting its focus entirely towards AI, as evidenced by its absence of gaming graphics cards at CES 2026 and the introduction of new AI products and architectures [2][10]. Group 1: AI Product Launches - NVIDIA unveiled the next-generation Rubin architecture GPU, which boasts inference and training performance that are 5 times and 3.5 times better than the Blackwell GB200, respectively [4][17]. - The company introduced five new product families targeting various AI applications, including the NVIDIA Nemotron for Agentic AI, NVIDIA Cosmos for physical AI, and NVIDIA Alpamayo for autonomous driving [6][8][39]. - The Vera Rubin NVL72 architecture was officially launched, featuring six core components designed to enhance AI data center capabilities [14][15]. Group 2: Performance Metrics - The Rubin GPU achieves an inference performance of 50 PFLOPS and a training performance of 35 PFLOPS under the NVFP4 data type, significantly surpassing its predecessor [17]. - Each Rubin GPU is equipped with 288GB of HBM4 memory and offers a bandwidth of 22 TB/s, supporting the high computational demands of modern AI models [18]. - The overall architecture of the Vera Rubin NVL72 can deliver 3.6 exaFLOPS of NVFP4 inference performance and 2.5 exaFLOPS of training performance [37]. Group 3: Networking and Connectivity - The introduction of NVLink 6 enhances interconnect bandwidth to 3.6 TB/s per GPU, with a total bandwidth of 260 TB/s across the entire NVL72 rack [20][21]. - The Vera CPU integrates 88 custom Arm cores and features a bandwidth of 1.8 TB/s for NVLink C2C interconnect, facilitating efficient communication between CPU and GPU [22]. Group 4: AI Model Developments - The Alpamayo model, a large-scale open-source visual-language-action model for autonomous driving, was launched with 10 billion parameters [41]. - The Nemotron series expanded to include specialized models for speech recognition, visual-language processing, and safety, enhancing AI applications across various sectors [49][51]. - The Cosmos model for robotics was upgraded to generate synthetic data that adheres to real-world physical laws, aiding in the development of AI agents [54][58]. Group 5: Industry Impact and Future Outlook - NVIDIA's comprehensive approach to AI, integrating models, data, and tools, is expected to strengthen its competitive edge and ecosystem lock-in [10]. - The company plans to begin mass production of the Vera Rubin NVL72 in the second half of 2026, indicating a strong commitment to advancing AI infrastructure [38].