AI Supply Chain: ASIC 2.0 – face-off on 3nm projectsAI
供应链：ASIC 2.0 - 面对3纳米项目的对决
Enlarged ASIC TAM: In our ASIC 2.0 global Insight we forecast AI ASIC TAM to grow from US$12bn in 2024 to US$30bn in 2027, amid strong competition from NVIDIA's AI GPU. Within ASIC, we are seeing several key 3nm projects finalizing:
扩大的ASIC总可用市场：在我们的ASIC 2.0全球洞察中，我们预测AI ASIC的总可用市场将从2024年的120亿美元增长到2027年的300亿美元，期间NVIDIA的AI GPU存在强烈的竞争。在ASIC领域，我们看到几个关键的3纳米项目正在完成：
AWS 3nm Trainium – Alchip vs. Marvell: We raise our CoWoS allocation for Marvell to ~70k, implying ~1.2mn units of Trainium chips, including Trainium2.5, which is migrating to HBM 12H for larger memory density. That upside should already be captured in Joe Moore’s recent price target hike. The enlarged Trainium2 market size (~30% rise, on our estimate) bodes well for the future Trainium3 market size for Alchip. In our view, Alchip is likely to handle the 3nm back-end design service given its good execution; Joe Moore also indicates that Marvell is getting involved in 3nm networking-related IP/design. We raise our 2026 revenue forecast for Altrip 23% to reflect the 3nm project upside. We think the next 2nm project (Trainium4) will see a new round of competition between Alchip and Marvell; the project award won’t be confirmed until 2H25. AWS 3nm Trainium – Alchip 对 Marvell：
我们将Marvell的CoWoS分配提升至约70k，意味着约120万颗Trainium芯片，包括正在向HBM 12H迁移以增大内存密度的Trainium2.5。这种上升应该已经在Joe Moore最近的目标价格提升中被体现。我们估计，扩大的Trainium2市场规模（约增长30%）对未来Alchip的Trainium3市场规模预示着好兆头。我们认为，鉴于Alchip良好的执行力，它很可能负责3nm后端设计服务；Joe Moore也指出，Marvell正在参与3nm网络相关的IP/设计。我们将Alchip 2026年的收入预测提高23%，以反映3nm项目的上行空间。我们认为，下一个2nm项目（Trainium4）将看到Alchip和Marvell之间的新一轮竞争；该项目的授予将不会在2H25之前确认。
Google 3nm TPU – MediaTek vs. Broadcom: The Google TPU v7 3nm project, amid some delay from the customer’s design schedule, is due to enter mass production in 2026. Its focus is more on training, given its high performance. We raise our MediaTek price target to NT$1,688, assuming a ~US$2.1bn revenue contribution in 2026 (ASP US$3,000 excluding HBM purchase). On the other hand, we believe Broadcom will extend another version of TPU v6 in 2026, (likely still at 5nm), which could be more efficient in inference.
Google 3nm TPU – MediaTek 对 Broadcom：Google TPU v7 3nm项目，因客户设计进度的一些延误，预计将在2026年开始大规模生产。其重点更多地放在训练上，鉴于其高性能。我们将MediaTek的目标价格提高到NT$1,688，假设2026年约US$2.1亿的收入贡献（ASP US$3,000，不包括HBM购买）。另一方面，我们认为Broadcom将在2026年推出另一个版本的TPU v6（可能仍然是5nm），这可能在推理方面更有效率。
Microsoft 3nm Maia200 – Marvell vs. GUC: We trim our GUC price target, as we believe Maia200 Enhancement by Marvell will be a larger opportunity than GUC’s Maia200 in 2026. But we still keep OW on GUC given its growing crypto mining chip demand.
Microsoft 3nm Maia200 – Marvell 对 GUC：我们降低了GUC的价格目标，因为我们认为Marvell的Maia200增强版将在2026年比GUC的Maia200提供更大的机会。但鉴于GUC的加密挖矿芯片需求增长，我们仍然保持对GUC的超配（OW）状态。
Meta 3nm AI networking – Broadcom vs Alchip: The MTIA v3 AI accelerator should migrate to HBM and CoWoS in 2026, and continue to use Andes's RISC-V IP. Our industry checks also suggest that Meta is customizing not only switch networking (link) but also RDMA-type networking chips for AI data centers. We believe Broadcom should be the first choice given its strong SerDes IP. We assume Alchip wins in future generations as our bull case.
Meta 3nm AI网络 – Broadcom 对 Alchip：MTIA v3 AI加速器应该在2026年迁移到HBM和CoWoS，并继续使用Andes的RISC-V IP。我们的行业检查还表明，Meta不仅在定制交换网络（链路），还在为AI数据中心定制RDMA类型的网络芯片。鉴于Broadcom强大的SerDes IP，我们认为Broadcom应该是首选。我们假设Alchip在未来几代中将作为我们的看涨案例获胜。
TSMC produces all above ASICs: TSMC will allocate 27% of CoWoS capacity for ASIC in 2025 with better margin. Total AI semi to account 30% of revenue in 2026. We OW TSMC/KYEC/MediaTek/Alchip/GUC/Andes as the key proxies for Greater China semis. Alchip is our new Top Pick given the biggest TP upside (45%), versus TSMC (30%), and Alchip's higher exposure to AI (75% of revenue).
台积电生产上述所有ASIC：台积电将在2025年为ASIC分配27%的CoWoS容量，以获得更好的利润率。2026年，AI半导体总收入将占30%。我们看好TSMC/KYEC/MediaTek/Alchip/GUC/Andes作为大中华区半导体的关键代理。鉴于Alchip具有最大的TP上升空间（45%），相较于TSMC（30%），以及Alchip对AI的更高曝光率（收入的75%），Alchip是我们的新顶级选择。
Price targets raised for Alchip (to NT$3,988), TSMC (to NT$1,388), and MediaTek (to NT$1,688): The increases mostly reflect greater AI ASIC and advanced packaging revenue assumptions in 2025 and 2026. Our new price targets imply P/Es of 42x, 23.5x, and 20x our 2025e EPS for Alchip, TSMC, and MediaTek, respectively. We think these levels are fair amid the trend of AI proliferation trend and market share gain from Asian ASIC providers.
为Alchip（提高至NT$3,988）、TSMC（提高至NT$1,388）和MediaTek（提高至NT$1,688）调高价格目标：这些提高主要反映了2025年和2026年AI ASIC及高级封装收入的增加假设。我们的新价格目标暗示Alchip、TSMC和MediaTek的2025年预期每股收益（EPS）的市盈率（P/E）分别为42倍、23.5倍和20倍。我们认为，在AI普及趋势和亚洲ASIC供应商市场份额增长的背景下，这些水平是公平的。
Of note, we move our Top Pick from TSMC to Alchip: This reflects the bigger share-price upside and stronger short-term catalyst – 3nm chip tape out in early 2025, although we also understand that TSMC is a much more liquid and less volatile stock.
值得注意的是，我们将我们的首选从台积电转移到Alchip：这反映了更大的股价上涨空间和更强的短期催化剂——3nm芯片在2025年初投产，尽管我们也理解台积电是一个流动性更强、波动性更小的股票。
Key downside risks for our new Top Pick Alchip (have been always the case):
1. CSP customers bring back chip design and production to internal teams
2. NVIDIA GPU gains further market share from ASIC in 2026
3. Alchip loses key projects, such as AWS’s next generation 2nm project
我们新的首选Alchip的主要下行风险（一直都是这样）：
1. CSP客户将芯片设计和生产内部化
2. NVIDIA GPU在2026年从ASIC进一步获得市场份额
3. Alchip失去关键项目，如AWS的下一代2nm项目
On the other hand, we trim our price target for GUC (to NT$1,600): This reflects some share loss on the Maia200 project. We remain OW in view of its crypto projects and slightly increasing AI business.
另一方面，我们调低了GUC的价格目标（至NT$1,600）：这反映了Maia200项目的一些份额损失。鉴于其加密项目和略有增长的AI业务，我们仍保持超配状态。
We would be more constructive on Egis/Novatek/Realtek if and when they win big projects.
如果Egis/Novatek/Realtek赢得大项目，我们会更具建设性。
大摩对ASIC芯片具体分析以及业绩展望，请关注本公众号并在后台发送“大摩ASIC芯片分析”获得。
专注分享财经信息的专业社区，喜欢请点击下方名片关注。
免责声明：本文材料整理于公司公告、新闻、机构公开研报及社交媒体网络等，对文中所涉专业及个股均不构成投资建议。市场有风险，投资需谨慎。图文如有侵权或违反信批规则请后台留言联系删除。
文章精彩，防止走失，可关注、转发、投币打赏。
