当算力焦虑从"地球烧到了太空"，市场早已开始寻找另一种缓解AI算力瓶颈的方式。 近日，片间光互连（OIO）创业公司光联芯科接连获得多家机构投资，成为国内光互连芯片赛道规模最 大的早期融资之一。光互连技术被业内视为突破当下算力瓶颈的关键路径，通过将芯片间短距互连从电 转向光，在传输能耗、带宽密度、延迟等核心指标上实现数量级提升。 中国算力底层技术架构正在发生深刻变革的同时，海外公司也在加快融资速度。据公开资料显示，美国 光互连公司Ayar Labs和光芯片创业公司Lightmatter也在去年年底完成了新一轮融资，其中不乏英伟达、 AMD、英特尔这样的投资者。 "在AI的发展中，买更多的GPU造更大的集群、追求更大的模型是一条路径，但边际效应是递减的。"光 联芯科CEO陈超在一次行业论坛中表示，英伟达从五年前就开始做光互连，头部企业很早就看到了电互 连的瓶颈，并有所准备。 从电到光 在他看来，国内数据中心的能力未来一要看算力，二看能耗。这两大能力不一定是在单芯片层面去解 决。如果利用光互连来赋能计算集群或者GPU公司，能够实现国内算力中心整体性能大于国外集群的目 标，并且能够降低包括芯片购买、芯片间的能耗等总体 ...

Group 1: Short-term Consensus - Trump's new tariff threat of 100% on all Chinese products starting November 1 has significantly increased market risk aversion, leading to declines in equities, industrial commodities, and cryptocurrencies, while gold and U.S. Treasuries rose[3] - The market consensus is that this is a maximum pressure strategy before the APEC summit at the end of October, indicating that the trend has not changed[4] - The overall economic impact is expected to be limited, with China's GDP growth target of 5% and a potential 5.2% growth in the first three quarters suggesting minimal macro policy adjustments before early November[4] Group 2: Long-term Competition - The trade war may signal the beginning of a decoupling between China and the U.S., entering a new cycle of fiscal expansion[4] - The shift from short-term TACO (Trump Always Chickens Out) trading to long-term confrontation is anticipated, with increased policy maneuverability in key sectors under the "high-quality development" strategy[5] - The potential for renewed TACO trading in early November could present opportunities in safe-haven assets, while risk assets may offer buying opportunities after adjustments[31]

Core Insights - Huawei's "Super Node + Cluster" strategy represents a critical response to external pressures and aims to establish a robust AI computing infrastructure in China [2][12][25] - The company emphasizes that AI competition fundamentally revolves around computing power, which is essential for advancing artificial intelligence capabilities [3][4] Group 1: Technological Innovations - The "Super Node" concept integrates multiple computing cards and machines into a single logical unit, enhancing overall efficiency and resource utilization [4][6] - Huawei's Atlas 950 and Atlas 960 SuperNodes support 8192 and 15488 Ascend cards respectively, significantly outperforming Nvidia's upcoming NVL144 in terms of scale and computing power [9][11] - The Lingqu interconnect protocol facilitates the collaboration of numerous computing resources, enabling a more efficient and scalable AI computing environment [13][15] Group 2: Strategic Positioning - Huawei's approach to AI computing is driven by the need to overcome limitations in semiconductor manufacturing due to sanctions, focusing on system architecture innovations [6][7] - The company aims to create a self-sufficient AI ecosystem that does not rely on Western technologies, addressing concerns over supply chain vulnerabilities [22][25] - Huawei's commitment to open-source initiatives is intended to foster collaboration within the industry, allowing for broader access to AI capabilities [19][20] Group 3: Market Implications - The deployment of AI computing resources through cloud platforms allows small and medium enterprises to access advanced AI models, democratizing AI technology across various sectors [23] - Huawei's long-term vision includes building a comprehensive AI ecosystem that supports sustainable growth and innovation within the Chinese AI industry [17][25] - The anticipated release of the Atlas 960 SuperNode and the Ascend 970 chip by 2028 is expected to enhance China's competitive position in the global AI landscape [25][26]

Core Viewpoint - Huawei's advancements in AI and computing capabilities, particularly through its Ascend chip series and supernode technology, position it as a formidable competitor in the global computing landscape, especially against NVIDIA [1][2][3]. Group 1: Huawei's AI and Computing Strategy - Huawei's strategy focuses on "supernodes + clusters" to overcome limitations in chip manufacturing processes and to provide sustainable AI computing power for China [2][4][6]. - The company has unveiled a roadmap for its AI chips, including the Ascend 950PR, 950DT, 960, and 970, with a planned release schedule that aims for a doubling of computing power approximately every year [8][9]. - Huawei's CloudMatrix 384 supernode is highlighted as the largest commercial supernode in China, capable of integrating 384 Ascend computing cards, thus achieving the highest effective computing power globally [4][5][9]. Group 2: Competitive Positioning - Huawei's Ascend 950 supernode is projected to maintain its status as the world's strongest computing supernode for several years, significantly outperforming NVIDIA's upcoming products in various metrics, including total computing power and memory capacity [9][10]. - The company emphasizes that its innovations are driven by necessity due to external pressures, leading to breakthroughs in computing architecture that redefine traditional models [5][12]. Group 3: Ecosystem and Future Outlook - Huawei aims to build a robust ecosystem around its hardware, promoting open-source software to attract developers and enhance hardware sales [11][12]. - The company is committed to developing its own AI ecosystem independent of Western technologies, with a focus on long-term positioning in the AI infrastructure market rather than direct monetization of large models [12][13].

Core Strategy - Huawei's core strategy is defined as "super nodes + clusters," which is central to its approach in the AI and computing landscape [1][2][10]. AI and Computing Capabilities - Huawei has developed a robust AI computing capability, claiming to have built the world's strongest computing cluster, with plans to launch new super nodes, including CloudMatrix 384 and Atlas 950 SuperPoD, in the coming years [2][12][16]. - The CloudMatrix 384 super node can support 384 Ascend computing cards, making it the largest commercial super node in China and one of the largest globally [10][12]. Competitive Positioning - Huawei's AI chips, including Ascend 910, are positioned to compete directly with NVIDIA's offerings, with Huawei asserting that its products will maintain a competitive edge in the market [9][16]. - The Atlas 950 super node is projected to outperform NVIDIA's upcoming NVL144 in several key metrics, including scale and total computing power [16][18]. Innovation Under Constraints - The company has been forced to innovate due to external pressures, particularly U.S. sanctions, which have limited its access to advanced chip manufacturing technologies [3][9][10]. - Huawei's approach to overcoming these challenges involves a shift from traditional computing architectures to a new "equal interconnection architecture," which enhances performance and scalability [11][19]. Future Product Roadmap - Huawei has outlined a clear product roadmap for its AI and computing chips, with multiple new products scheduled for release between 2026 and 2028, aiming for a doubling of computing power approximately every year [15][16]. - The company plans to introduce new series of AI chips, including Ascend 950PR, Ascend 950DT, and Ascend 960, alongside advancements in its Kunpeng CPU line [15][16]. Ecosystem Development - Huawei emphasizes the importance of building a robust ecosystem around its hardware, including open-source software initiatives to attract developers and enhance the usability of its products [20][21]. - The company aims to establish a self-sufficient ecosystem that does not rely on Western technologies, particularly in the AI domain, to ensure long-term sustainability and growth [21][22].

Core Viewpoint - Huawei's rotating chairman Xu Zhijun announced that the Ascend chips will evolve at a pace of "almost doubling computing power every year" over the next three years [4][5]. Group 1: Ascend Chip Development - Huawei has planned three series of Ascend chips: Ascend 950, Ascend 960, and Ascend 970, with more specific chips in development [5][6]. - The Ascend 950 series includes Ascend 950PR and Ascend 950DT, featuring significant improvements such as support for low-precision formats, increased vector computing power, and a bandwidth of 2TB/s, which is 2.5 times that of the Ascend 910C [5][7]. - The Ascend 960 is expected to launch in Q4 2027, with specifications that double those of the Ascend 950, including support for a proprietary HiF4 data format [6][7]. - The Ascend 970 is still under discussion, with preliminary plans indicating a doubling of computing power and bandwidth compared to the Ascend 960 [7]. Group 2: Super Nodes and Clusters - Huawei will launch three super node products: Atlas 950, Atlas 960, and TaiShan 950, along with a new interconnect protocol called Lingqu [9][10]. - Xu emphasized that super nodes are becoming the dominant product form for AI infrastructure, functioning as a single logical computer composed of multiple physical machines [9]. - The Atlas 950 SuperCluster will be 2.5 times larger than the current largest cluster, xAI Colossus, and will have 1.3 times its computing power, making it the strongest computing cluster globally [10]. - By Q4 2027, Huawei plans to introduce the Atlas 960 SuperCluster, further increasing the scale to a million card level [10].

图源：观察者网 他还公布了以昇腾950芯片为基础的新型超节点，将成为全球最强超节点，甚至比英伟达将在2027年推 出的NVL576系统更强。 （文/观察者网 吕栋 编辑/张广凯） 9月18日，在华为全连接大会2025上，华为轮值董事长徐直军宣布了多颗昇腾系列芯片和演进路线，包 括昇腾950系列、昇腾960系列和昇腾970系列。 图源：观察者网 他提到，华为自研了低成本HBM，将以一年一次算力翻倍的进度推进，支持FP8等更多精度格式，更大 的互联带宽。 徐直军表示，有了昇腾芯片为基础，就能满足客户的算力需求，超节点将成为AI基础设施建设的新常 态。 不仅如此，以昇腾960为基础的超节点，也将在2027年四季度上市，持续提供充沛算力。 徐直军还提到，华为还将更新通用计算鲲鹏处理器，包括鲲鹏950系列和鲲鹏960系列，以及以这些芯片 为基础的超节点。 图源：观察者网 他还宣布了面向超节点的互联协议"灵衢"，把更多计算资源连接在一起，以昇腾950为基础可以组成超 50万卡集群，以昇腾960为基础甚至可以组成超过99万卡的集群。 图源：观察者网 徐直军提到的这些芯片和技术，将在2026-2027年相继上市。 "我们单颗 ...