Core Viewpoint - Nvidia's new Blackwell Ultra and Rubin chips are set to lead the next wave of AI investment, focusing on both generative and inferential AI, with significant improvements in memory bandwidth and processing capabilities [1][3][11]. Summary by Sections Blackwell Ultra and Rubin Chips - The Blackwell Ultra chips, set to be released in 2024, feature a memory bandwidth increase from 192GB to 288GB, enabling the handling of larger AI models and more efficient processing of intensive workloads [1]. - The Rubin chips, expected in 2026, will replace Blackwell and consist of a custom Arm CPU named Vera, designed for parallel processing, allowing for a modular design that accommodates multiple chip types for specific workloads [2]. Performance Enhancements - Rubin is projected to achieve AI inference speeds of 50 petaflops, approximately 2.5 times faster than Blackwell, with plans for a combined Rubin GPU to reach 100 petaflops and 1 terabyte of onboard memory by 2027 [3]. - The Blackwell Ultra and Rubin chips are designed to enhance inference capabilities, addressing the growing demand for complex, dynamic AI models that require real-time adjustments [3][4]. Market Position and Competition - Nvidia currently dominates the AI accelerator market, supporting about 90% of data center workloads, which solidifies its high valuation and status as one of the world's most valuable companies [5]. - The Blackwell Ultra system claims to provide 11 times faster AI inference speed and 7 times higher throughput compared to the previous generation, indicating a strong competitive edge [5]. Industry Dynamics - Competitors like Google and Amazon are developing their own chips tailored to their ecosystems, not necessarily to compete directly with Nvidia but to optimize their resource utilization [7]. - AMD is making strides with its Instinct MI300X chip, which is being integrated into data centers by companies like Microsoft and Meta, indicating a growing competitive landscape [7][8]. Future Outlook - Nvidia's new chips are expected to open up new applications for AI, transitioning the company from a hardware manufacturer to a critical interface between hardware and software [6]. - Despite the emergence of competitors, Nvidia's software ecosystem, including CUDA, remains a significant barrier for rivals, as replicating this infrastructure is a daunting task [9]. Investment Sentiment - Analysts predict that the AI data center market could reach trillions of dollars, with Nvidia positioned at the center of this growth, supported by high gross margins of approximately 73-75% [13]. - The recent stock price decline of Nvidia is viewed as irrational, with the current market conditions presenting a potential buying opportunity for investors [13][14].

【申万海外科技】英伟达 FY25Q4：Blackwell 产能爬坡逐渐恢复 ＃业绩整体超预期、但相对此前季度超预期幅度有所收窄，计算业务超预期、网络低于预期。 整体营收 393 亿美元（YoY+78%，预期 382 亿），Non-Gaap 净利润 221 亿（YoY+72%，预期 210 亿），其中数据中心计算收入 356 亿（YoY+93%，预期 341 亿）、网络收入 30 亿（YoY+9%， 预期 36 亿）。 ＃业绩指引： FY26Q1 营收 430 亿（预期 422 亿），并预计随着 Blackwell 完全恢复供应链 爬坡后再 FY26 年底达到 75%毛利率水平（FY25Q4 为 73.5%）。 ＃Blackwell 出货进展逐渐恢复，FY25Q4 Blackwell 系列实现 110 亿美元营收，GB300 系列即 将陆续上线，下一代 Blackwell Ultra 将于 25H2 发布。 ＃本季度大型 CSP（微软、谷歌、亚马逊、Oracle）占公司数据中心收入的近一半，消费者 互联网企业收入同比增长 3 倍、包括 xAI、META 等。 | 农. 央节人财务调要 (2023-2-21) | ...