证券研究报告 科技 再谈 NV 的下一个 Mellanox：Groq LPU 的整合 华泰研究 科技 增持 (维持) | 何翩翩 | 研究员 | | --- | --- | | SAC No. S0570523020002 | purdyho@htsc.com | | SFC No. ASI353 | +(852) 3658 6000 | | 易楚妍 | 联系人 | | SAC No. S0570124070123 | yichuyan@htsc.com | | SFC No. BXH065 | +(86) 21 2897 2228 | | 韩冬冰* | 联系人 | | SAC No. S0570125070150 | handongbing@htsc.com +(86) 21 2897 2228 | 重点推荐 本报告延续我们 2026 年 1 月 12 日发布的《英伟达吸收 Groq 定义 AI 下半 场》观点。彼时我们指出，英伟达整合 Groq 的战略，与其 2020 年收购 Mellanox 一脉相承，核心在于吸收人才及将领先的底层 IP 内生化，以补齐 架构层面的结构性短板。尽管市场普遍预期英伟达可能 ...

In my last article on Nvidia Corporation ( NVDA ), published in November 2025, I analyzed the company’s Q3 report and explored the key takeaways, with a special focus on the rising competitionAn independent investor in the Indian and US equity markets with a CFA Charter and a PhD in Finance from University of Durham, U.K. I hold an Honorary Associate Professor in Finance and Corporate Governance title at Brunel University London. I have a YouTube and a Podcast channel, titled The Stock Doctor,' where I disc ...

In my last article on Nvidia Corporation ( NVDA ), published in November 2025, I analyzed the company’s Q3 report and explored the key takeaways, with a special focus on the rising competitionAn independent investor in the Indian and US equity markets with a CFA Charter and a PhD in Finance from University of Durham, U.K. I hold an Honorary Associate Professor in Finance and Corporate Governance title at Brunel University London. I have a YouTube and a Podcast channel, titled The Stock Doctor,' where I disc ...

Summary of NVIDIA 2026 Conference Call Company Overview - **Company**: NVIDIA (NasdaqGS: NVDA) - **Conference Date**: March 04, 2026 - **Key Financials**: - Revenue: $70 billion - Net Income: $46 billion [39] Core Industry Insights - **AI and Compute**: - The last two years have seen three significant inflection points in AI technology, leading to a surge in demand for compute resources [24][31]. - Companies like OpenAI and Anthropic are compute-limited, meaning their revenue potential is directly tied to their compute capacity [34]. - **Token Economy**: - The concept of AI factories is introduced, where data centers are viewed as factories producing tokens, which are monetizable [33]. - The relationship between compute and revenue is emphasized, with a direct correlation established: more compute leads to higher revenues [34][35]. - **Market Dynamics**: - The software industry is predicted to evolve into a token-driven economy, where companies will need to rent or produce tokens to drive their operations [42][78]. - The entire IT industry is expected to consume vast amounts of tokens, significantly increasing its market size [44]. Strategic Developments - **Ecosystem Building**: - NVIDIA is focused on building a robust ecosystem around its CUDA platform, investing in AI-native companies to extend its reach [55]. - Recent investments include a $30 billion commitment to OpenAI and a $10 billion investment in Anthropic, aimed at expanding their compute capacities [56][57]. - **Technological Advancements**: - NVIDIA has pioneered advancements in AI, including physical AI and autonomous systems, positioning itself at the forefront of these technologies [63][64]. - The company has developed a full-stack approach to computing, integrating hardware and software to optimize performance [22][35]. Financial Considerations - **Funding and CapEx**: - The need for substantial capital expenditure (CapEx) to support the growing compute demands is highlighted, with a focus on the importance of making informed decisions regarding infrastructure investments [32][38]. - NVIDIA's strong balance sheet is seen as a strategic advantage in securing supply chains and supporting its growth initiatives [49]. Additional Insights - **Future Predictions**: - The speaker predicts that compute will become essential for all companies, equating it to GDP growth for nations [76]. - The transition to agentic software is anticipated, where every software company will leverage AI models to enhance their operations [41][43]. - **Constraints and Opportunities**: - Current constraints in memory, power, and infrastructure are viewed positively, as they force companies to make optimal choices in their technology investments [46][50]. This summary encapsulates the key points discussed during the NVIDIA conference, focusing on the company's strategic direction, industry insights, and financial implications.

Huang’s Keynote, 1,000+ Sessions and Breakthroughs Across the AI Stack Headline the World’s Leading AI Conference NVIDIA CEO Jensen Huang and Global Technology Leaders to Showcase Age of AI at GTC 2026 Huang’s keynote, 1,000+ sessions and breakthroughs across the AI stack headline the world’s keading AI conference. SANTA CLARA, Calif., March 03, 2026 (GLOBE NEWSWIRE) -- NVIDIA today announced that GTC, the world’s premier conference on AI and accelerated computing, will take place March 16-19 this year i ...

公众号记得加星标⭐️，第一时间看推送不会错过。 听 到 "25 年 前 " 这 个 词 很 容 易 忽 略 上 下 文 ， 所 以 我 们 来 谈 谈 GeForce 3 发 布 时 的 背 景 。 那 时 还 没 有 Steam，也没有iPhone，更没有YouTube。"GPU"的概念当时还很新颖，GeForce 256仅仅在15个月 前才正式发布。没错，我们从GeForce 256到GeForce 2再到GeForce 3，仅仅用了15个月的时间。那 时候科技发展速度更快；CPU从1GHz到2GHz的升级也差不多用了同样的时间。 GeForce 3 于 2001 年 2 月发布，是图形处理器发展史上的一个关键转折点。它是首款真正意义上具 备可编程性的 GPU，因为它支持DirectX 8.0 像素着色器和顶点着色器。这意味着图形程序员现在可 以编写在 GPU 上运行的程序。 你看，GeForce 3 拥有强大的 DirectX 8 图形功能，但它的原始光栅化性能（填充率）与 GeForce 2 Pro 完全相同。它在 DirectX 7 及更低版本游戏中的唯一真正优势在于其"光速内存架构"，这是一种 当 ...

瓦西什塔似乎赞同这种观点。他在MWC前夕接受采访时表示："如果你考虑一下整个技术栈，5G Advanced和6G的开源使得整个技术栈的代码都能被开发者使用。我认为，对于一些公司，尤其是规 模较小的公司来说，要达到这种程度的灵活性一直是一个相当大的挑战。" 他说，如果6G广泛采用开 源技术，那么拥有新的波束成形算法（一项5G时代的技术）的开发者，理论上就可以将其集成到更 大的平台中。 英伟达已经拥有一个名为 Aerial 的开源无线接入网 (RAN) 参考平台。这使得像 DeepSig 这样的开 发 者 能 够 实 现 Vasishta 所 描 述 的 功 能 ， 即 在 Aerial 协 议 栈 中 插 入 AI 原 生 波 形 。 有 趣 的 是 ， DeepSig 是为 OCUDU 构建参考平台的两家小型公司之一，另一家是总部位于爱尔兰的 SRS。英伟 达 AI-RAN 方案的一部分内容是，将人类经过数十年改进的 RAN 算法替换为 AI，以提高频谱效 率。 然而，Aerial 仍然要求开发者主要在 CUDA 框架内工作。CUDA 是英伟达的旗舰软件平台，通常被 视为该公司的防御屏障。在无线接入网 (RA ...

但 FPGA、eFPGA 和 DSP 的设计既复杂又耗时。Arteris 产品管理与营销副总裁安迪・奈廷格尔 （Andy Nightingale）称："FPGA 的应用场景理应比仅用于原型验证或特定功能更加广泛。在降低 内存和 I/O 瓶颈方面，它们非常理想。但对 FPGA 进行编程仍然相当复杂。与在 GPU 上运行软件完 成类似任务相比，编写 FPGA 代码需要具备 RTL 设计能力。" 虽然 FPGA 工程师已经对比特流的输入输出方式进行了优化，但这需要一套不同的软件栈来管理。 Baya Systems 首 席 商 务 官 南 丹 ・ 纳 亚 姆 帕 利 （ Nandan Nayampally ） 指 出 ： " 赛 灵 思 （ 现 已 并 入 AMD）和阿尔特拉等公司都构建了核心 CPU 集群，使其 FPGA 架构具备更强的可编程能力。他们 试图解决部分编程难题，但要打造一套能同时适配 GPU、CPU 和 FPGA 的通用方案非常困难。软件 栈种类越多，快速推进的难度就越大。" 公众号记得加星标⭐️，第一时间看推送不会错过。 人工智能正开始介入可编程逻辑的设计与管理工作，可用于简化并加速设计流程中的部分 ...

英伟达的黄仁勋为何只持有3.5%的股份，如果他不减持，那他不就是世界首富了。不过这个假设是不存在的。微软1992年上市的市值，比尔盖茨拥有48%的 股份，之后他不断减持，目前持股不到1%。他减持的股份，换来了大约500亿美元的资金。 有人就说了，微软现在市值几万亿美元，如果比尔盖茨不减持，他就是第一个身家破1万亿的地球人了。可是这个假设是没意义的。如果比尔盖茨不减持， 机构投资者怎么进入微软，这些机构投资者不拿到微软的股份，谁在资本市场去给微软抬轿子。 同样的道理也发生在英伟达、苹果这些公司身上。创始人跟管理层的持股都是非常少，而且股权极度分散。 对于黄仁勋来说，也是如此。他现在只有3.5%的股权，看似如果他一路不减持，他现在可以有几万亿美元的财富，可是如果他不减持，又怎么可能把英伟 达做起来呢？ 1993年黄仁勋开始创业，当时他跟两个合伙人一起合作，各自投资200美元，每一个人都是占股三分之一。也就是说从一开始，黄仁勋就不是持股100%。 因为黄仁勋等人缺乏资金，因此为了公司的发展，他们在1993年就拿了红杉资本等机构的2000万美元融资。这个时候黄仁勋的持股就被稀释到30%左右。2 年之后，英伟达面临破产 ...

Core Insights - Nvidia reported a remarkable 73% year-over-year revenue increase and a 75% rise in net profits for Q4 FY26, highlighting strong financial performance [2] - Despite the focus on new Blackwell chip rollouts, demand for older Ampere (A100) chips remains significant, contributing over $20 billion to Nvidia's quarterly data center revenue [3][5] Demand for Legacy Chips - Approximately one-third of Nvidia's $62.3 billion data center revenue is driven by older architectures, including the A100 and Hopper chips, indicating a critical reliance on these legacy products [3] - The A100 chip is priced around $10,000 on the secondary market, making it more affordable compared to the newer Blackwell GPUs, which can cost up to $50,000 [5] Software Ecosystem and Customer Lock-In - Nvidia's proprietary CUDA software ecosystem enhances customer retention, as it ties developers to Nvidia's architecture, making transitions to competitors costly and complex [4][9] - The integration of CUDA with low-level GPU programming and high-performance libraries creates a strong vendor lock-in effect, which is difficult for competitors to overcome [9][10] Competitive Landscape and Future Risks - While Nvidia currently leads in training stack and developer ecosystem, there is a potential risk as inference becomes more dominant in AI computing, which may lead to increased competition from custom ASICs developed by companies like Alphabet and Amazon [11][12] - If inference accelerates faster than expected, Nvidia's market share in data center expenditures could decline, impacting margins as clients may opt for lower-cost, task-specific chips [12][13]