老黄演讲后，网上很多博主都发了关于GTC的内容，但是绝大多数都是新闻性质的，他们只讲了黄仁勋都说了啥。这篇文章我们结合 NVIDIA 的技术规 划，来聊一聊网上可能没有的分析和GTC的核心看点。对于万亿营收和CPO的信息，都是大家知道的了，这篇文章就不再赘述了。 1、CPX的黯然退场 在上个月下旬的时候，网上还依然传着很多关于 CPX 要使用 HBM 的传言。我当时就听到消息说 CPX 要取消，我当时还不太信，但事实证明确实是被 LPU 取代了。 那为什么会取代呢？就是因为英伟达的路线转换，他们要从 prefill 加速切换到 推理加速。 这些内容和观点，我们之前在星球中都讲过。 2、谁是这次 GTC 大会之后最大的受益方？ 那显然是三星。因为 LPU 是找三星独家代工，采用的是三星的 N4 工艺。这意味着三星不仅是英伟达全层级存储的核心供应商，现在更独揽了 LPU 的 代工大单。 在Rubin上，三星的综合价值量已经超过台积电了，因为台积电只负责代工加封装。 原文链接：https://globalsemiresearch.substack.com/p/nvidia-gtc-2026-is-samsung-t ...

LPU 会带来哪些增量？20260311 摘要 LPU 采用片内集成大量 SRAM 替代片外 HBM，通过近存计算与编译器 预调度消除延迟瓶颈，核心解决 AI 推理中 Decode 环节的内存密集型瓶 颈。 LPU vs GPU/TPU：LPU 追求极致确定性，更接近特定模型定制 ASIC；GPU 侧重通用性依赖 HBM；TPU 侧重矩阵乘法效率，LPU 在低 时延推理上具代差优势。 推理市场呈现"PD 分离"趋势：Prefill 环节为计算密集型，适用高算力 CPX；Decode 环节为内存密集型，LPU 的 SRAM 架构可显著降低 KVCache 读取时延。 NVIDIA 存储布局多元化：SRAM（收购 Groq/LPU）攻低延迟场景， HBM4 攻高性能训练，GDDR 攻高性价比计算，SSD 探索独立 KV 缓存 机柜。 Groq 产业化提速：2025 年三星代工订单预计从 0.9 万片增至 1.5 万片； 技术增量指向 SRAM 3D 堆叠、背部供电 PCB、高速 Switch 及液冷方 案。 LPU 局限性：SRAM 成本远高于 DRAM，大模型推理需多芯片堆叠推高 初始成本，且软件栈灵活性不足， ...

Investment Rating - The industry investment rating is "Overweight" [5]. Core Insights - The GTC 2026 conference is expected to focus on AI inference evolving into system-level infrastructure, with key components including Rubin Ultra and Feynman architectures [2][3]. - The year 2026 is anticipated to be a pivotal year for Agentic AI, with significant developments in CPO and LPU technologies [2]. - The integration of Groq into NVIDIA's ecosystem is expected to enhance LPU capabilities, transitioning from a standalone LPX rack to a more integrated approach within NVIDIA's GPU roadmap [3][4]. Summary by Sections GTC 2026 Conference Highlights - NVIDIA's GTC 2026 will feature a keynote by CEO Jensen Huang, focusing on AI infrastructure advancements [2]. - Key technologies discussed will include Rubin Ultra, Feynman, CPX, and LPU, with expectations for significant updates on these platforms [2]. Rubin Ultra and Feynman Architecture - Rubin Ultra is expected to integrate 144 GPUs in a Kyber rack design, with a total power consumption of approximately 600 kW [2]. - Feynman may adopt TSMC's A16 process technology and is projected for a 2028 launch, with potential outsourcing of I/O die to Intel [2][3]. LPU and LPX Developments - The LPX rack is seen as a transitional solution, with LPU capabilities expected to be integrated into NVIDIA's GPU roadmap starting from the Feynman architecture [3]. - The LPU version of the LPX rack is being evaluated for a 256 LPU configuration, indicating a significant increase from the initial 64 LPU version [3]. CPO and Optical Interconnects - CPO and optical interconnects are anticipated to be central themes at GTC 2026, with a focus on the evolution from Scale-Out to Scale-Up architectures [4]. - The introduction of Scale-Up CPO switches is expected to complement the Rubin Ultra architecture, with significant bandwidth capabilities [4][10].

①和②共识很强，③讲的少一些，但NV的护城河也离不开消费者洞察。 NV的产品需求帮我们确立了很多需求变化：大模型对高显存的需求（HBM）、训练对集群的需求（十万卡集 群）、开发者编译对大节点的需求（NVL72）、汽车机器人对端侧大算力 3.8周观点：NV如何应对Agentic AI新增的算力需求？ NV的护城河：①技术引领；②长期算力执着；③消费者洞察。 3.8周观点：NV如何应对Agentic AI新增的算力需求？ NV的护城河：①技术引领；②长期算力执着；③消费者洞察。 ①和②共识很强，③讲的少一些，但NV的护城河也离不开消费者洞察。 NV的产品需求帮我们确立了很多需求变化：大模型对高显存的需求（HBM）、训练对集群的需求（十万卡集 群）、开发者编译对大节点的需求（NVL72）、汽车机器人对端侧大算力的需求（Orin/Thor）、上下文窗口对内 存的需求（SSD边柜）、推理对PD分离的需求（CPX）等。 NV推一个新系列产品，不是单纯为了塞一个很牛的新技术给客户，而是因为存在现有产品无法精准覆盖的新需 求，进而新开一个产品序列。 用深刻的消费者洞察，在全球最领先的工具库里定义产品，长期坚守算力而"等"来了 ...

Investment Rating - The report indicates a positive investment outlook for the electronic industry, suggesting an "Overweight" rating, indicating that the industry is expected to outperform the overall market [22]. Core Insights - The electronic sector has shown resilience, with significant stock price increases for key companies such as Mingyang Circuit (43.54%) and Ruikeda (30.71%) during the review period [4][5]. - Nvidia's acquisition of Groq's technology IP for approximately $20 billion is expected to enhance its AI inference capabilities, integrating Groq's Language Processing Unit (LPU) technology into its product lineup [7][12]. - Shenghe Jingwei, a leader in 2.5D packaging, is set to raise 4.8 billion yuan for advanced packaging projects, with a projected revenue growth from 1.633 billion yuan in 2022 to 6.521 billion yuan in 2025 [13]. Summary by Sections Market Performance - From February 23 to February 27, the A-share electronic index rose by 4.07%, with the Shanghai Composite Index increasing by 1.98% and the Shenzhen Composite Index by 3.10% [4][5]. - Key stocks in the electronic sector experienced substantial gains, with Mingyang Circuit leading at 43.54% [5][6]. Nvidia and Groq Technology - Nvidia's integration of Groq's LPU technology is anticipated to provide significant advantages in AI inference, leveraging a unique architecture that emphasizes static scheduling and deterministic execution [8][12]. - The LPU's design allows for low-latency token generation, although it requires a large network scale due to its limited on-chip SRAM capacity [8][9]. Shenghe Jingwei's Market Position - Shenghe Jingwei is recognized as the leading company in China's 2.5D packaging market, holding an 85% market share, and is expected to continue its growth trajectory with increasing revenues projected through 2025 [13]. - The company aims to utilize the funds raised from its IPO to enhance its 3D integration and packaging capabilities [13].

Investment Rating - The report rates the electronic industry as "Overweight," indicating that the industry is expected to outperform the overall market [4]. Core Insights - The report highlights that Nvidia plans to integrate Groq technology to enhance its inference capabilities, with Groq's LPU (Language Processing Unit) showcasing significant performance advantages over traditional GPU/TPU architectures [8][9]. - The report also notes that Shenghe Jingwei, a leader in 2.5D packaging, has passed regulatory approval for fundraising, aiming to raise 4.8 billion yuan for advanced packaging projects [14]. Summary by Sections Market Overview - From February 23 to February 27, the A-share electronic index rose, with the Shanghai Composite Index increasing by 1.98%, the Shenzhen Composite Index by 3.10%, and the ChiNext Index by 1.05%. The semiconductor index in the U.S. fell by 1.96% [5][6]. Nvidia and Groq Technology - Nvidia acquired Groq's technology IP for approximately $20 billion, with key personnel joining Nvidia. Groq's LPU is designed for AI inference, utilizing a 14nm process with 230MB on-chip SRAM and 80TB/s bandwidth, demonstrating superior performance and energy efficiency [8][9]. - The LPU's low latency is achieved through pre-compilation and deterministic execution, eliminating uncertainties in dynamic instruction scheduling [9][10]. Shenghe Jingwei's Market Position - Shenghe Jingwei is the leading company in China's 2.5D packaging market, holding an 85% market share. The company is projected to achieve revenues of 1.633 billion yuan in 2022, growing to 6.521 billion yuan by 2025 [14]. Investment Recommendations - The report suggests focusing on domestic computing power, AI storage, and semiconductor equipment parts, highlighting companies such as SMIC, Hua Hong, and others as potential investment opportunities [4].

Core Insights - The article highlights the escalating competition in the AI infrastructure sector, marked by significant investments from major tech companies like Nvidia and Alibaba, indicating a strategic shift towards building powerful computing capabilities for AI development [1][2][5]. Group 1: Major Investments and Strategic Moves - Nvidia and OpenAI recently announced a monumental $100 billion deal to develop next-generation AI supercomputing clusters [1]. - Alibaba has committed to investing 380 billion RMB (approximately $53 billion) in AI infrastructure, joining the ranks of other tech giants like OpenAI, Google, and Meta in the global "computing power war" [2][3]. - The article emphasizes that advanced algorithm models are essential for entering the race towards Artificial General Intelligence (AGI) and Superintelligence (ASI), with robust computing infrastructure being the core battlefield [5]. Group 2: Strategic Challenges in Building Computing Empires - The construction of a successful computing empire requires more than just financial investment; it demands foresight, engineering excellence, innovative system architecture, and a strong developer ecosystem [6]. - The challenges faced by industry players are universal, as they all strive to establish their own "computing barriers" in this competitive landscape [7]. Group 3: Nvidia's Strategic Partnerships - Nvidia's recent $5 billion investment in Intel to co-develop customized data center and PC products has generated significant industry buzz, reflecting a dramatic shift from past rivalries to collaboration [10]. - This partnership is expected to enhance product competitiveness, particularly in the laptop market, while revitalizing Intel's position in the industry [10]. Group 4: GPU Market Dynamics - The GPU market has experienced dramatic fluctuations, likened to a "drug trade," with supply shortages and price wars affecting availability and pricing strategies [12]. - New entrants in the cloud service market have intensified competition, leading to a complex landscape where acquiring GPUs for large-scale deployment remains a significant challenge [12]. Group 5: Oracle's Rise in Cloud Services - Oracle has emerged as a dark horse in the cloud services market, leveraging its substantial balance sheet to support large-scale computing orders for clients like OpenAI [13]. - Its flexible hardware strategy allows Oracle to deploy the most effective technology combinations, enhancing its competitive edge [13]. Group 6: Amazon AWS's Recovery Strategy - Amazon AWS is experiencing a resurgence after a growth slowdown, driven by its vast data center resources and the provision of massive GPU and custom chip capabilities to major clients [14]. - Despite challenges with its custom chip Trainium, AWS is adapting its infrastructure to meet the demands of AI workloads [15]. Group 7: New AI Hardware Opportunities and Challenges - The introduction of Nvidia's Blackwell architecture marks a new era in AI hardware, presenting both performance advancements and new challenges regarding cost, reliability, and system architecture [16]. - The GB200 architecture presents a performance paradox, where its deployment costs are higher, but the performance gains are highly workload-dependent [17]. Group 8: Nvidia's Competitive Edge - Nvidia's success is attributed to its visionary leadership, particularly Jensen Huang's bold decision-making and execution capabilities, which have allowed the company to maintain a significant competitive advantage [22][24]. - The company's ability to deliver new chip designs successfully on the first attempt is a testament to its engineering prowess and operational efficiency [26]. Group 9: Future Considerations for Nvidia - Nvidia faces the challenge of effectively utilizing its substantial cash flow for future investments, with options including infrastructure development and AI factory expansions [27].

Financial Performance - The company reported the strongest revenue performance in its history for both Q4 and the full year, with Q4 revenue of $227.1 million, a 14% increase year-over-year, and full-year revenue of $879.7 million, a 16% increase [7][9][10] - Adjusted gross profit for Q4 was $83.9 million, up 15% from the prior year, and adjusted EBITDA was $51.7 million, reflecting a 16% improvement [9][10][28] - For the full year, adjusted gross profit increased by 19% to $328.1 million, and adjusted EBITDA grew by 21% to $204.3 million [10][28] Business Segment Performance - The SMB segment generated Q4 revenue of $155.7 million, an 11.1% increase year-over-year, with adjusted gross profit of $32 million, up 0.4% [18][20] - The B2B segment saw revenue of $23.7 million, a 10.9% increase, with adjusted gross profit rising 24% to $6.4 million [22][23] - The Enterprise segment reported Q4 revenue of $48.7 million, a 27% increase, with adjusted gross profit also increasing by 27% to $45.6 million [24][25] Market Data - The company processed over $130 billion in annual transaction volume and managed over $1.2 billion in average daily account balances [7] - The percentage of adjusted gross profit from recurring revenues reached over 63% in Q4, nearly doubling since early 2022 [17] Company Strategy and Industry Competition - The company aims to sustain momentum in its acquiring, payables, and Enterprise segments, projecting 10% to 14% revenue growth for 2025 [8][36] - The focus remains on innovation in SaaS payments and banking solutions, with a commitment to refining customer experiences [12][41] - The company is positioned as an undervalued player compared to industry peers, with a strategy to leverage its public currency for potential acquisitions [18][42] Management Commentary on Operating Environment and Future Outlook - Management expressed confidence in achieving revenue growth despite potential headwinds from lower interest rates and a challenging macroeconomic environment [8][36] - The company is actively working to address a material weakness in internal controls over financial reporting, which did not affect the financial results [34][35] Other Important Information - The company redeemed its preferred stock, resulting in a fourth-quarter preferred dividend of $2.65 million [31] - A $10 million prepayment on the term loan was made post-quarter end, with plans to continue reducing the cost of capital [32][36] Q&A Session Summary Question: Capital allocation strategy between debt pay down and increased spending - Management will continue to evaluate the best use of capital, balancing debt pay down with potential M&A opportunities [50] Question: Impact of interest rate cuts on guidance - Management has factored in expected interest rate cuts into their 2025 guidance, taking a conservative approach [54][56] Question: Impact of tariffs on business - Management noted that tariffs have not significantly impacted the acquiring side, but there is increased demand for B2B strategies due to constrained working capital [67] Question: Explanation of gross margin decline in consumer payments - Management clarified that organic margins have increased, and the decline is due to residual purchases running off [73][78] Question: Magnitude of CapEx to OpEx shift - The expected impact from the CapEx to OpEx shift in 2025 is approximately $4 million [110]