以下文章来源于More Than Semi ，作者Nico More Than Semi . More Than SEMI 半导体行业研究 这篇文章，结合业内的一个专家纪要，聊一下超节点。 超节点领域的合作与方案 在超节点领域，下游CSP大厂主要找服务器供应商定制整机柜产品。服务器厂商，包括国产GPU厂商， 也在规划自己的超节点机柜，但以NV方案为主。在国内，主流超节点方案包括腾讯的ETH-X、NV的 NVL72或其变形、华为昇腾的CM384、阿里的磐久，这些已在市场推广或者有客户购买。字节在规划 大模型的以太网创新方案，以博通Tomahawk为主，还未推广。海光的方案多为传统刀片式服务器，与 NV和华为的以太网或NVLink互联不同，目前海光无类似创新方案。 互联网大厂自研方案与华为384、海光96卡、摩尔跨柜方案本质不同在于应用偏向：CM384偏训练和大 模型计算；腾讯ETH-X偏推理；字节大模型整机柜偏训练和高性能计算。服务器厂商在合作中需具备交 换机和服务器开发能力，与博通、NV、国产GPU厂商有关系。在AI服务器8卡、16卡、32卡方案有市场 占有度，才能进入CSP选型。目前入选供应商有华勤、华三、 ...

昨天赶在阿里云栖大会的最后一天，特意从上海赶去杭州看了阿里的超节点，在现场拍了一个视频，但现场环境比较杂乱，人非常多。 | 42 | | | --- | --- | | ব J | | | 40 | | | 30 | | | 38 | | | 37 | ipmi0002 | | 36 | ipmi0001 | | 35 | | | 34 | 1U Power Shelf 33kW | | 33 | 1U Power Shelf 33kW | | 32 | 1U Compute Tray | | 31 | 1U Compute Tray | | 30 | 1U Compute Tray | | 29 | 1U Compute Tray | | 28 | 1U Compute Tray | | 27 | 1U Compute Tray | | 26 | 1U Compute Tray | | 25 | 1U Compute Tray | | 24 | 1U Compute Tray | | 23 | 1U Compute Tray | | 22 | 1U Non - Scalable NvSwitch5 Tray ...

Summary of Key Points from the Conference Call Industry Overview - **Industry**: Global Technology, specifically focusing on AI Supply Chain and semiconductor industry - **Key Players**: TSMC, Nvidia, Broadcom, AMD, MediaTek, Alchip, Aspeed, Advantest, KYEC, and others Core Insights and Arguments 1. **Cloud Capex Growth**: Morgan Stanley expects 2026 cloud capex to increase by 31% year-over-year (Y/Y) to US$582 billion, significantly higher than the consensus estimate of 16% Y/Y growth [2][10][67] 2. **AI Server Capex**: The implied AI server capex is projected to grow approximately 70% Y/Y in 2026, driven by an increase in AI server capex mix [2][11] 3. **TSMC CoWoS Capacity**: TSMC's 2026 CoWoS capacity is estimated to remain unchanged at 93k wafers per month, which is expected to support the anticipated AI capex growth [1][4][27] 4. **Hyperscalers' Capex**: The top 11 global hyperscalers are tracking a cash capex of US$445 billion for 2025, reflecting a 56% Y/Y growth, with major contributions from Microsoft, Amazon, and Alphabet [9][68] 5. **AI Chip Demand**: Strong demand for AI semiconductors is expected to persist, with Morgan Stanley maintaining an "Overweight" rating on key semiconductor companies [2][11] Additional Important Insights 1. **China's AI Chip Supply**: There are ongoing developments in China's AI chip supply, with Nvidia placing new wafer orders at TSMC for H20 chips, adding 200k units to the previous production of 1 million units [8] 2. **Nvidia's CoWoS Consumption**: Nvidia's CoWoS consumption assumption has been raised from 580k to 595k units for 2026, indicating strong demand for their chips [22] 3. **Market Dynamics**: The report highlights the competitive landscape, noting that Huawei is developing its own AI chips to compete with Nvidia, showcasing the CM384 server rack prototype [8][41] 4. **Depreciation Trends**: Depreciation as a percentage of total expenses for data center customers is expected to rise, reflecting increased investments in data centers [60] 5. **AI Inference Demand**: Monthly tokens processed by major cloud service providers indicate a growing demand for AI inference, with Google processing over 980 trillion tokens in July 2025, doubling from May [15] Conclusion The conference call emphasizes a robust outlook for the AI semiconductor industry, driven by significant increases in cloud capex and AI server investments. TSMC's capacity plans and the competitive dynamics in the AI chip market, particularly with developments in China, are critical factors to monitor.