液冷产业相关介绍我们在前期文章反复提及，在此不过多赘述。本文章更多叙述2026年液 冷行业供需两侧逻辑/数据及国内厂商的出海机会展望。 千亿的液冷赛道 英伟达Rubin的液冷新方案？ GB200 GB300液冷价值量拆解 需求侧 1、随着单AI算力卡及机柜功耗越来越高，风冷无法满足相应算力卡及机柜的散热要求，必须要上液 冷方案才能解决相应散热问题。（单柜功耗＞35-40KW即无法上风冷） 2、北美缺电严重，液冷能显著降低PUE，大幅节约单位算力密度总能耗，缓解缺电导致的数据中心 项目推迟。（液冷方案PUE＜1.2），不要小看这一点点的降低，数据中心的电费是非常夸张的，这 也是为什么国内很多数据中心都建在内蒙古、新疆等地的原因。 3、北美不少数据中心项目离经济中心或居民区域相对较近，液冷相对风冷散热方案能显著降低数据 中心机房噪音，促使数据中心项目落地进程加快，减少资金垫付财务成本 目前英伟达GB200及以上更高级机柜系列方案全面运用液冷散热方案；谷歌的TPU V7 及以上更高级 机柜系列方案全面运用液冷散热方案（此前TPU V3有尝试运用液冷方案，但当下量不大）；Meta的 MTIA V2及以上更高级机柜系列 ...

Equity Research e] Sector Report | Technology Dec 8, 2025 AI Networking: The Key Move in 2027 to be CPO in NVIDIA's Scale Up Alicia Xia Jeff Pu, CFA Evan Lee aliciaxia@gfgroup.com.hk jeffpu@gfgroup.com.hk evanlee@gfgroup.com.hk What's new Following Google's adoption of OCS optical interconnection, we believe that NVIDIA (NVDA US Buy) may incorporate CPO (rack to rack optical interconnection within Kyber) into the scale-up networking of one version of its 2027 Rubin Ultra architecture. As for NVIDIA's OIO so ...

Summary of ASIC Industry Update Industry Overview - The document provides an update on the ASIC (Application-Specific Integrated Circuit) projects of major North American hyperscaler companies, including AWS, Microsoft, Meta, Google, OpenAI, Apple, and TikTok [2][3] Key Companies and Their ASIC Projects AWS (Amazon Web Services) - **Tranium 2 Chip**: Expected to reach its end phase in Q4 2025, with a transitional chip, Tranium 2.5, to be produced in Q4 2025 and Q1 2026. Marvell is expected to ship approximately 200,000 units per quarter [4][3] - **Tranium 3 Chip**: Forecasted production volume of around 2.5 million units, with potential allocation of up to 500,000 units to Marvell if Tranium 2.5 production is successful [8][9] - **Tranium 4 Chip**: Designed by Annapurna and Alchip, expected to start mass production in Q4 2027 [9][10] Microsoft - **Cobalt 200 CPU and MAIA 200 Sphinx**: Designed by GUC, with MAIA 300 Griffin facing challenges in its development with Marvell. Microsoft may shift to Broadcom if confidence in Marvell wanes [14][16] - **MAIA 200 and MAIA 300**: Part of the second-generation ASIC accelerator series, with the contract with Marvell expiring in H1 2026 [15][16] Meta - **ASIC Roadmap**: Includes multiple generations of chips, with the first-generation inference chip, Artemis, already in mass production. The second-generation training chip, Athena, is set for Q4 2023, and the third-generation chip, Iris, is planned for Q3 2024 [17][18] - **Arke Chip**: A simplified inference-only chip designed by Broadcom and Marvell, expected to help Meta keep pace with NVIDIA's chip iterations [19][20] Google - **TPU Development**: The first-generation ASIC Server CPU, Axion, is designed by Marvell, while the second-generation, Tamar, is designed by GUC. Google expects to produce about 4 million TPUs in 2026, with significant internal use [22][24] - **Demand Surge for Optical Modules**: Due to the increase in TPU production, demand for 1.6T optical modules is expected to rise dramatically from 3 million units in 2025 to 20 million in 2026 [25][26] OpenAI - **Titan 1 and Titan 2 Chips**: Broadcom is developing these chips, with expected shipments of 300,000 units in 2026 and at least 600,000 units in 2027 [28][29] - **Collaboration with ARM**: OpenAI is also working with ARM on ASIC projects, indicating a dual approach to chip development [30][31] Apple - **ASIC Projects**: Apple is customizing two ASIC chips, with mass production not expected before 2027 [32][33] TikTok - **Neptune Chip**: After negotiations, TikTok is expected to resume mass production of its ASIC chip in Q1 2026, with an anticipated production volume of 500,000 units [34][35] GUC (Global Unichip Corp) - **Controversial Position**: GUC is involved in the production of Google's Tamar CPU but is also engaged in more profitable projects like Tesla's AI5 chip, which could generate significant revenue in 2027 [41][43] Additional Insights - The document highlights the competitive landscape among major players in the ASIC market, with companies like Marvell, GUC, and Broadcom playing crucial roles in the design and production of these chips [41][42] - The anticipated growth in demand for ASIC chips, particularly in the context of AI and machine learning applications, suggests a robust market outlook for the coming years [25][26] This summary encapsulates the key developments and projections within the ASIC industry, focusing on the major players and their respective projects.