Marvell Technology Group Ltd | North America Quick Thoughts on NVDA/ MRVL Partnership NVIDIA and Marvell partner on NVLink and silicon photonics, with a $2bn investment, underscoring how severe networking bottlenecks have become and highlighting the strategic positioning of Marvell's portfolio as AI scaling drives greater reliance on advanced interconnect. What Happened: NVIDIA and Marvell announced a partnership to integrate NVIDIA's NVLink ecosystem with Marvell's XPU and scale-up networking portfolio. Ma ...

Flash | 31 Mar 2026 07:17:14 ET │ 8 pages Taiwan Semiconductors and Electronic Components & Equipment AI chip structural upgrade cycle intact CITI'S TAKE Companies Mentioned: ASE Technology Holding (3711.TW; NT$328.5; 1; 31 Mar 26; 13:30) | King Yuan Electronics Co (2449.TW; NT$261.0; 1; 31 Mar 26; 13:30) | NVIDIA Corp (NVDA.O; US$165.17; 1; 30 Mar 26; 16:00) | TSMC (2330.TW; NT$1760.0; 1; 31 Mar 26; 13:30) | Unimicron (3037.TW; NT$444.5; 1; 31 Mar 26; 13:30) | WinWay Technology (6515.TW; NT$7150.0; 1H; 31 ...

MARCH 2026 See Disclosure Appendix of this report for important disclosures and analyst certifications ARTIFICIAL INTELLIGENCE: THE AI INFRASTRUCTURE VALUE CHAIN Our fact base and framework for comparing fundamental upside and identifying winners across the AI value chain. The AI boom versus bubble debate has raged on since 2023. The range of outcomes could scarcely be wider, from advanced superintelligence to a crash of epic scale. While it is clearly difficult to pinpoint the exact trajectory, we see AI s ...

Dissecting Nvidia Blackwell - Tensor Cores, PTX Instructions, SASS, Floorsweep, Yield 剖析英伟达 Blackwell 架构——张量核心、PTX 指令、SASS、残次品利用、良率 Microbenchmarking, tcgen05, 2SM MMA, UMMA, TMA, LDGSTS, UBLKCP, Speed of Light, Distributed Shared Memory, GPC Floorsweeps, SM Yield 微基准测试、tcgen05、2SM MMA、UMMA、TMA、LDGSTS、UBLKCP、光速指标、 分布式共享内存、GPC 残次品利用、SM 良率 KIMBO CHEN AND DYLAN PATEL KIMBO CHEN 与 DYLAN PATEL APR 01, 2026 2026 年 4 月 1 日 ∙ PAID ∙ 付费内容 136 1 Share 分享 Nvidia's Datacenter Blackwell GPU (SM100) represents one of the ...

台积电调研：CPO 进展更新，新增设备需求，产能规划， 供应商格局 20230331 摘要 CPO 仍处于研发阶段，量产良率尚不明确，预计 Feynman 架构阶段 才会显著放量并导入 3D 堆叠技术。 NVIDIA 明确跳过 NPO 方案，坚定推动 CPO 以解决 GB200/300 铜缆 供应瓶颈及散热干扰问题。 台积电主导 CPO 前端晶圆级工艺，SPIL 负责后端封装；CPO 产能准备 预计 2028 年启动，2029 年放量。 Onto Innovation 2026 年产能已售罄，订单排至 2027Q1，预计 2026 年营收增长 38%-50%。 Onto G5 设备已获海力士订单，正处于台积电与美光验证期，旨在切入 CoWoS 与 HBM 检测增量市场。 三星计划在 HBM4 激进导入 Hybrid Bonding 技术，而海力士与美光 因成本考量暂无跟进计划。 LPU 项目由三星代工，但 SRAM 堆叠封装由台积电负责，目前正在权 衡 3nm/4nm/5nm 制程选择。 Q&A 目前的良率水平仍处于研发（R&D）阶段，并非量产良率。在样品阶段，良率 有时可以达到 50%至 60%，但量产时的 ...

OCS 交换机行业进展更新 20230331 摘要 OCS 应用场景从 TPU 柜间连接拓展至跨数据中心 Scale- across，2026 年 Oracle 已下单，微软、Meta、英伟达进入样机测试 阶段。 市场规模由通道数驱动，2026 年主流产品从 128 通道升级至 300+ 通道，预计 2028 年总需求量达 7-8 万台，年增速有望翻倍。 MEMS 方案最成熟，液晶方案（Coherent/微软）2026 年起量且寿命 更长，光波导方案时延短但目前仅能产 32 通道，产业链配套尚不成熟。 OCS 整机售价约 500 美元/通道，BOM 成本约 200 美元/通道；当前 供给瓶颈在于整机装配效率低，预计 2026Q3 随供应链爬坡进入大规模 量产。 腾景科技为液晶方案核心供应商，提供晶体材料及准直器阵列（占 BOM 40%），预计 2026 年相关收入超 2 亿元。 德科立主攻光波导整机代工并送样谷歌/英伟达；光库科技具备代工经验； 凌云光与英唐智控分别布局压电陶瓷与收购技术团队。 英伟达未来架构（如 2028 年费曼代）Scale-up 互联是否采用 OCS 尚 无定论，若落地将成为产业趋势加速 ...

Core Insights - Chinese GPU and AI chip makers captured nearly 41% of China's AI accelerator server market in 2025, significantly reducing Nvidia's market dominance [2][5] - The shift is attributed to China's increasing caution about reliance on foreign chips, leading to a push for domestic alternatives amid U.S. export controls [3][5] Market Share Dynamics - Nvidia remains the market leader with approximately 2.2 million AI accelerator cards shipped, holding a 55% market share, but this represents a decline from its previous dominance [4][7] - Chinese vendors collectively shipped 1.65 million cards, marking a significant milestone in the market, with Huawei leading at around 812,000 cards, followed by Alibaba's T-Head with approximately 265,000 cards [5][6] Competitive Landscape - Baidu's Kunlunxin and Cambricon each shipped around 116,000 cards, ranking them jointly third among Chinese vendors [6] - Other Chinese vendors like Hygon, MetaX, and Iluvatar CoreX accounted for 5%, 4%, and 3% of total shipments, respectively [6] Government Initiatives - In 2025, the central government initiated a new wave of AI infrastructure spending, encouraging local governments to accelerate the establishment of intelligent computing centers with directives to prioritize domestic products [6]

Iran's Islamic Revolutionary Guard Corps (IRGC) has threatened attacks on a swath of U.S. tech companies with operations in the Middle East, including Nvidia, Apple, Microsoft and Google.The IRGC warned on Tuesday that 18 tech companies would be considered as "legitimate targets" in retaliation for U.S. and Israeli strikes on Iran.Attacks on those companies would begin from 8 p.m. on Wednesday, April 1, Tehran time (10:30 am ET), the IRGC said in a post on Telegram translated by Google, warning employees at ...

Light-duty vehicles travel over 3 trillion miles annually in the U.S., per the Bureau of Transportation Statistics. Ride-sharing services typically charge $1 to $2 per mile. So, even if autonomous driving technology cuts the price in half, robotaxis still represent a trillion-dollar market in the U.S. alone.In the fourth quarter, hedge fund billionaire Ken Griffin bought shares of Nvidia (NVDA +5.56%) and Amazon (AMZN +3.51%), two stocks at the center of that opportunity. In fact, Nvidia and Amazon were the ...

自动驾驶作为物理AI的起点，正在开始"二级跳"。 英伟达所定义的"物理AI"（Physical AI） 则提出了一种不同于传统学习驱动架构的技术方向，其核心是 理解现实世界中的因果关系与物理规律，通过仿真、合成数据生成和端到端模型来训练AI在物理环境 中的感知、推理与行动能力，在此基础上实现与现实世界的交互。这一变化并非单点技术升级，而是面 向真实世界应用的系统级重构，尤其体现在对极端和长尾场景的处理能力上。 映射到自动驾驶领域，这意味着系统不再仅根据感知结果或模型输出直接生成控制指令，而是需要在复 杂交通环境中识别潜在风险、评估多种可能情境，并推演下一步行为。当面对施工区域、非常规交通行 为或未被充分覆盖的场景时，系统需要以接近人类驾驶员的方式进行推理，而不是完全依赖大量特定场 景的数据训练。 从CES到GTC，英伟达正以肉眼可见的速度持续加码自动驾驶。在今年的GTC大会上，黄仁勋强调最多 的一句话："下一波人工智能的浪潮是物理AI。AI将理解物理世界的规律，而自动驾驶汽车，就是我们 目前能看到的最庞大、最成熟的具身智能机器人。" 为此，英伟达不仅推出DRIVE AV全栈自动驾驶平台，更一口气拉上比亚迪、吉 ...