Shares of Arm (ARM) rocketed more than 15% higher in early trading Wednesday, after the company unveiled its first production data center processor: the Arm AGI CPU (central processing unit). Arm has traditionally licensed its intellectual property to other companies to develop their own chips, including Apple and Nvidia (NVDA), which uses Arm’s capabilities in its Grace and Vera CPUs. Graphics processing units, or GPUs, have dominated data centers thanks to their ability to train and run AI models. Bu ...

For years, Arm (ARM) has played a key role in the development of processors for everything from the iPhone to data center chips. But now the company is expanding its reach with the debut of its first production data center processor: the Arm AGI CPU (central processing unit). Arm has traditionally licensed its intellectual property to other companies to develop their own chips, including Nvidia (NVDA), which uses Arm’s capabilities in its Grace and Vera CPUs. Graphics processing units, or GPUs, have do ...

通信&人工智能"周周谈" — GTC 大会总结 20260318 摘要 NVIDIA 定义数据中心为"Token 工厂"，核心逻辑从技术演进转向每 瓦电力 Token 产出最大化，驱动 AI 服务实现非线性定价弹性。 官方指引 2027 年底 Rubin 与 Blackwell 平台累计收入达 1 万亿美元， 且未计入 GraceCPU、GrokLPU 及 RubinUltra 等潜在增量。 AI 应用 Token 消耗量与年化收入呈指数级增长，Azure AI 业务及 OpenAI 收入在两个月内接近翻倍，NIM 被定位为驱动企业级应用的新 操作系统。 Rubin 平台采用极致协同设计，整合 7 种芯片与 5 种机柜，其中 CPU 机 架为今年纯新增量，专为 Agent 时代高单线程性能需求设计。 硬件规格大幅升级：计算托盘 PCB 层数需达 50 层以上，Rubin 平台 100%采用液冷散热，机架安装时间由 2 天缩短至 2 小时。 混合推理架构实现任务拆分：RubinGPU 负责高吞吐预填充，GrokLPU 负责低延迟逻辑任务，后者延迟较传统方案降低约 10 倍。 连接方案采取差异化策略：主流 72 ...

This isn’t wild conjecture. Just last month, Nvidia inked a massive multiyear deal with Meta Platforms ( NASDAQ:META ) to deploy millions of its Grace CPUs in standalone configurations, alongside Blackwell and upcoming Rubin GPUs. The partnership explicitly includes a roadmap for Vera CPU-only servers in 2027, signaling Meta’s bet on discrete CPUs for scaling AI workloads.Speculation has run wild in the weeks leading up to GTC. Some analysts expect details on the Rubin GPU architecture (successor to Blackwe ...

Core Insights - Nvidia is set to reveal new details about its agentic-optimized CPUs at the upcoming GTC conference, highlighting the growing importance of CPUs in AI workflows as they become the bottleneck in processing [2][4] - The company has seen a significant increase in demand for its CPUs, with a multiyear deal with Meta marking the first large-scale deployment of its Grace CPUs, and plans for Vera CPUs in 2027 [3][4] - The CPU market is projected to grow from $27 billion in 2025 to $60 billion by 2030, driven by the shift towards agentic AI applications [4][10] Nvidia's CPU Strategy - Nvidia's CPUs, including the Grace and upcoming Vera, are designed specifically for data processing and agentic AI workflows, differing from traditional general-purpose CPUs [12][15] - The company utilizes Arm architecture for its CPUs, contrasting with the x86 architecture used by competitors Intel and AMD, which positions Nvidia's products as optimized for AI workloads [15][20] - Nvidia's CPUs are expected to enhance the performance of its GPUs, with a focus on single-threaded performance rather than maximizing core count [13][14] Market Dynamics - The CPU market is experiencing a "quiet supply crisis," with leading providers like AMD and Intel facing supply shortages and increased delivery lead times [10][11] - Nvidia's robust supply chain has managed to meet demand effectively, as many of its CPUs are sold alongside GPUs in integrated systems [12] - The server CPU market share as of late 2025 shows Intel leading at 60%, followed by AMD at 24.3%, and Nvidia at 6.2%, with a growing presence of in-house Arm-based CPUs from major hyperscalers [18] Competitive Landscape - Nvidia's strategy remains platform agnostic, supporting both Arm and x86 architectures, and has opened its NVLink technology for third-party licensing to facilitate integration with other CPUs [19][20] - The company is adapting to the competitive landscape by expanding its product offerings beyond GPUs to include CPUs and specialized hardware, addressing a diverse range of workloads [20]

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 ...

科技巨头Meta与芯片大厂AMD正式达成重磅合作！ 根据双方最新协议，Meta将采用AMD定制版InstinctGPU，产品预计2026年下半年开始出货。AMD还向 Meta授予附带业绩条件的认股权证，后者最多可认购1.6亿股AMD普通股。 截止发稿，受此利好消息推动，AMD股价大涨近7%，报209.93美元；而Meta股价小幅走弱，微跌 0.23%，报635.79美元。 订单价值高达600亿美元 当地时间2月24日，AMD与Meta正式宣布达成五年战略合作协议，围绕下一代AI基础设施展开深度绑 定。根据协议，AMD将为Meta提供总算力达6吉瓦的定制版Instinct GPU，这笔订单价值高达600亿美 元。 该GPU基于MI450架构打造，专为Meta业务负载优化，首批1吉瓦规模的部署预计于2026年下半年启动 出货。 除硬件采购外，协议还包含一项基于业绩表现的股权激励条款。AMD向Meta授予了股票认股权证，允 许后者以每股0.01美元的行权价分批购买至多1.6亿股AMD普通股。若双方达成特定的技术与商业里程 碑，且AMD股价触及600美元目标位，Meta最终可能获得AMD约10%的股份，成为其核心股东 ...

1）豆包：除夕AI互动19亿次，Seedance2．0为春晚提供技术支持；2）豆包App暂停视频通话功能；3）反超千 问、蚂蚁阿福，豆包App Store排名第一；4）2月12日凌晨，智谱发布新一代旗舰模型 GLM-5。 5）Meta加码英伟达，未来数年部署数百万颗芯片，首次采用Grace CPU。 < 国产算力（6）：算力紧缺+主流大模型适配国产芯片印证算力需求旺盛【东北计算机】 1）豆包：除夕AI互动19亿次，Seedance2．0为春晚提供技术支持；2）豆包App暂停视频通话功能；3）反超千 问、蚂蚁阿福，豆包App Store排名第一；4）2月12日凌晨，智谱发布新一代旗舰模型 GLM-5。 国产算力（6）：算力紧缺+主流大模型适配国产芯片印证算力需求旺盛【东北计算机】 5）Meta加码英伟达，未来数年部署数百万颗芯片，首次采用Grace CPU。 #春节流量峰值验证算力瓶颈，国产算力迎来规模化商用拐点。 2026 年春节期间 AI 应用场景的集中爆发，直接凸显出算力资源的供给缺口，国内 AI 产业对算力的刚性需求进一 步释放。 1）字节跳动旗下豆包 App 因新春用户流量大幅攀升出现算力承压，于 2 ...

公众号记得加星标⭐️，第一时间看推送不会错过。 在人工智能的推动下，英伟达的GPU广为人知。但从最近的消息看来，这家GPU巨头正在大举进 军CPU。 日前，英伟达与 Meta 签署了一项扩大的多年数据中心协议，根据该协议，这家芯片制造商将向这家 社交媒体巨头提供数百万个 Blackwell 和 Rubin GPU。虽然这无疑是这则新闻中最引人注目的部 分。 但两家公司表示，根据协议，Meta 还将在其数据中心推出仅搭载 Nvidia Grace CPU 的服务器，这 是该芯片的首次大规模部署。Grace 是英伟达与两块 Blackwell 或两块 Blackwell Ultra GPU 搭配 使用的处理器，用于组成其 GB200 和 GB300 AI 超级芯片。 从这次他们与Meta的合作，我们可以看到他们对于服务器CPU方面的决心。而且，英伟达推出仅支 持 Grace 芯片的服务器之际，正值超大规模数据中心运营商越来越多地寻求利用传统 CPU 芯片来帮 助支持一些 AI 推理和智能 AI 应用，从而抓住市场对传统 CPU 日益增长的需求。 由此可以断定，英伟达正在以另一种方式进军服务器CPU。 强势杀入服务 ...

Core Insights - NVIDIA Corporation is gaining attention on Wall Street due to an expanded partnership with Meta, with a Buy rating and a price target of $250.00 from Stifel analyst Ruben Roy [1][8] Group 1: Partnership Details - The strategic partnership between NVIDIA and Meta is multiyear and multigenerational, focusing on on-premises, cloud, and AI infrastructure, including large-scale deployment of NVIDIA CPUs, GPUs, and Spectrum-X Ethernet networking [2] - Meta will adopt NVIDIA's rack-scale Confidential Computing as a foundational feature for the upcoming Vera Rubin platform [2] - This collaboration marks the first large-scale deployment of Grace CPU only, supported by infrastructure co-design and software optimization investments from both companies [3] Group 2: Implementation and Future Prospects - Meta will initially implement NVIDIA's Confidential Computing for WhatsApp private processing, with potential expansion to other use cases [4] - The expanded collaboration reinforces the structural alignment between NVIDIA and Meta, validating NVIDIA's 'extreme co-design' initiative, transitioning from a chip provider to a full AI platform solutions provider [5] - Despite competition from hyperscalers investing in internal silicon projects, NVIDIA is expected to maintain effective competition through a full-stack architecture approach [5] Group 3: Company Overview - NVIDIA specializes in AI-driven solutions, providing platforms for data centers, self-driving cars, robotics, and cloud services [6]