CoWoS七年翻倍 + 存储vs半导估值激辩 —— 买方视角的AI仓位地图

2026-06-27 周六 (ET) · 主帖 12 / 回复 33
以下部分引用 AI 总结,现阶段 AI 仍然有幻觉。请以内容的原文为准。
一些重点
  1. CoWoS需求大升级2027全球CoWoS wafer同比+93%,AMD用量4倍跳,MTK v9成关键变量
  2. AVGO vs MTKMTK吃份额但AVGO美元价值更高,不是'赢家通吃'
  3. AMD被低估?MI455故事定价不足,Lisa执行信誉与T-Glass供应是市场担忧
  4. 存储vs半导估值撕裂多空激辩:LTA是结构护城河还是价格封顶?
  5. Anthropic $3TGavin Baker喊3万亿估值,IPO可能成流动性抽水
  6. 中国替代新热度2026年半导体设备国产化率目标40%,AMEC/NAURA最确定
  7. USG投资前沿实验室国家安全逻辑可能推动政府注资,解释OpenAI推迟IPO
  8. Cloudflare架构重组CEO说AI杀'测量者',保建设者和销售者
  9. 地缘情绪余波中东风险延续,政权被'8头蛇'逻辑绑架

1 CoWoS格局重构,ASIC供应链博弈升级

CoWoS 2027全景 多·供给瓶颈

2027年CoWoS wafer需求翻倍,AMD用量超越Broadcom MS预测全球CoWoS wafer从2026年139.4万片增至2027年269.4万片,同比+93%[1]NVDA仍占45%,但AMD从13万片跳升至53万片(+4x),份额升至20%高于AVGO的18%[1]

  • AMD MI455是2027 GPU主力 + Venice CPU放量: MS预计TSMC CoWoS-L booking约24万片;Venice CPU非TSMC CoWoS booking约27万片→约675万颗CPU产能[1]
  • CoWoS-R = AI CPU交易,市场尚未定价: Vera CPU + AMD Venice CPU让CoWoS-R成新瓶颈;若agentic AI推高CPU需求,这一层还没被充分建模[1]
  • GUC四年4倍增长: MS预计GUC CoWoS booking从2026年1.4万片增至2027年6万片(+4x YoY),绝对规模小但增速极强[1]

CoWoS是AI硬件稀缺性的放大器,AMD+MTK的份额上升比NVDA的绝对值更重要

AVGO vs MTK 争论·TPU份额

MediaTekGoogle份额,但Broadcom美元价值更高 群内分析对比:MTK v9 2nm+HBM4e是重估原因,但AVGO 2027年CoWoS wafer 2.7倍于MTK,大die高价值[1]

  • 鱼名越取越离谱,但die效率差距清楚: Broadcom Sunfish每wafer 12颗,MTK Zebrafish每wafer 20颗→AVGO更大die更少颗,但每颗价值更高[1]
  • MTK v9 2nm订单:2027年通过Intel EMIB-T出货 约40万颗2nm TPU v9预计通过Intel EMIB-T出货[1]
  • AVGO不止是Google proxy,CLSA客户清单扩容: ASIC客户覆盖Google、AnthropicOpenAIMetaByteDance等,市场过度聚焦Google份额流失[1]Jalapeno首次流片即A0,2026年底可部署→可能推动sellside模型上调收入。

对PM = 押注MTK是押注技术代际跃迁,押注AVGO是押注多客户ASIC广度

⬇️ AMD GPU风险折价 观察·低估争议

MI455故事定价不足,但AMD CEO执行信誉+供应能力受质疑 群内多方认为如果MS数字属实AMD太便宜[2];反驳点在于:(1)AMD CEO缺乏NVIDIA CEO级别的AI可信度(2)T-Glass等关键物料供应可能比NVDA难获取[3]。关注标的:AMD、GUC、MTK[4]

  • 市场为什么折扣AMD的GPU路线指引?: Arun Dhar追问。回应指出:不仅信誉gap,也是供应链执行的信心问题;AMD在关键材料如T-Glass的供应获取上可能比NVIDIA更难[5][3]

2 存储 vs 半导估值大辩论 —— 一场关于护城河与倍数的正面对决

估值撕裂:10x vs 40x 争论·估值框架

估值框架之争:LTA是护城河还是价格封顶? davii发问:为什么熊专打MU/SNDK,不打PE 50x的AMAT[6] 引发全场最激烈辩论。

  • 熊方逻辑1:政治化风险压制估值 FZ:若USG施压存储器扩产,MU面临产能提前释放;中国DRAM ETF若允许直接投资CXMT/YMTC→市场给存储更低倍数[7]
  • 熊方逻辑2:双杀情景——存储最脆弱 ALC:对AI trade最大的两个风险是hyperscaler降capex+FOMC加息,存储器两者都推,最脆弱[8]
  • 多方反击:半导护城河真的深吗? Arun Dhar长篇论述:半导是做设备集成(买工具),存储器是做原子级材料科学→可模仿性不同。存储器盈利波动是半导的3倍,必须靠供给纪律维持信任;半导则是利基垄断。同GICS并不意味着同倍数[9]
  • davii反击:LTA改变周期性 + PayPal/Visa类比 过去周期的波动不适用于今天,LTA存在意味着未来更可预测。PayPal倍数跟存储器差不多,Visa是AMAT的一半。中国能淹存储器为什么不能淹半导?ASML无敌,但NAURA/AMEC/Skyverse呢?[10]
  • Arun Dhar还击:LTA是价格封顶,不是结构性护城河 内部有价格上限、地板重谈、take-or-pay退出条款,最好的时候没用,最差的时候最先崩。并指出davii的Visa类比'litigating the wrong half of the analogy'[11]
  • Terafab作为摇摆因子: John补刀:本周All In提到Terafab专注制造存储芯片,可能改变产能格局[12]

辩论核心:LTA是结构护城河还是价格封顶,这决定了估值的归属

3 AI前沿估值 + 中国替代 + 其他信号

中国半导体国产化 多·替代主线

设备国产化2026年达40%,AMEC/NAURA最确定 CLSA预计国产化率从2022年18%→2025-26年30-40%[1]AI芯片本地化从2025年30-40%→2030年60-70%→需要扩建先进逻辑、DRAMNANDHBM产能。

  • DRAM+HBM双扩产:CXMT产能至少翻倍 CXMT 2025年产能280kwpm,国内DRAM未来3-5年至少翻倍。HBM设备强度比标准DRAM高30-50%[1]
  • YMTC产能翻倍 + 蚀刻沉积占capex 55-60%: YMTC 2025年160kwpm,NAND至少翻倍。3D NAND中蚀刻+沉积占capex 55-60%[1]
  • 直接受益标的:AMEC + NAURA CLSA预计AMEC 2026年订单+48% YoY,NAURA存储客户订单+50%+ YoY。这是中国AI+存储+替代三合一的直接交易[1]

Anthropic $3T + Claude Tag 观察·估值争议

3万亿估值下,Anthropic的IPO是流动性抽水 All In上Gavin Baker称Anthropic值3万亿美元。AMZN持股已稀释至约15%以下[13]SmutCap再次确认2-3万亿估值,但提示token使用量增长放缓+IPO可能成为流动性抽水[14]

  • Claude Tag vs Slackbot:Anthropic与Salesforce的微妙竞争 Anthropic推出Claude Tag(Slack AI助手),与Salesforce自家的Slackbot功能重叠,而Slackbot恰好也是用Anthropic技术。Salesforce员工表示困惑[15]
  • Mike戏称:ZM是买入标的 若Anthropic估值3万亿且AMZN持股稀释至15%以下,按AMZN 2.5万亿市值计算占比不大[16]

USG投资前沿实验室? 观察·政策

国家安全逻辑可能推动美国政府注资前沿AI实验室 TheLastWhiteRhino提出:开源不确定性+数据中心成本上升→USG可能将前沿实验室视为国家安全必需,这也解释了OpenAI推迟IPO的传闻[17][18]

Cloudflare架构重组 观察·企业结构

AI不是消灭工作,是消灭'测量者'角色 Cloudflare裁员1100人后工程团队反增45%(1308→1894)。CEO Prince框架:企业分建造者/销售者/测量者,AI消除第三类[19]

地缘情绪余波 空·宏观

中东地缘风险延续,政权被'8头蛇'逻辑绑架 H-Cap指出政权看不上财富收益(石油美元化、解冻资产),SmutCap补充:可能最终招来更严重的军事反击[20][21]

数据新鲜度群聊 08-18 ✓卖方 08-18 ✓Wrap 断档50天 ✕News 断档None天 ✕X 断档47天 ✕页面生成 08-19
📊 今日群聊 ATOMS
23 atoms 🟦 fact 3 🟥 take 20 15 / ▼ 5 / ◆ 0
实体 (今日): AMD (4), Anthropic (2), CoWoS-R (1), GUC (1), Broadcom vs MediaTek (1)
主题 (本周): 估值 (23), 竞争 (16), 定价 (11), 技术路线 (9), 资本开支 (8)
Δ 今日变化
  • AWS forecast 本月到期: 「Q2 营业利润指引需达 $14-15bn 高端才能推动股价」(2026-04-27 由 chat 提出)
  • Meta forecast 本月到期: 「META 指引需达 $61bn (28% yoy) 以支撑 AI 投资」(2026-04-27 由 chat 提出)
  • 🌱 新叙事: 出货预测 (今日 6 条, 库内首见)
  • 🌱 新叙事: 替代主线 (今日 4 条, 库内首见)
📜 今日原 atoms (按 entity 折叠展开)
AMD · 4 atoms
  • 🟥 MI455 CoWoS-L booking 预期 2027 · 产能/出货预测
    MS sees MI455 as the key 2027 GPU product, with TSMC CoWoS-L booking rising to around 240k wafers.
  • 🟥 Venice CPU 非 TSMC CoWoS booking 预期 2027 · 产能/出货预测
    But the other big driver is Venice CPU, where non-TSMC CoWoS booking rises to around 270k wafers, implying around 6.75m units of CPU production.
  • 🟥 市场低估 MI455 故事 · 估值争议
    Jukan asks: How much of the MI455 story do you think is currently priced into AMD’s stock? Terminal Junkie: Not much.
  • 🟥 CEO 执行信誉与供应担忧 · 管理层信誉/供应链
    I think part of it is that Lisa hasn’t built the same level of market credibility/accountability around AI guidance that Jensen has. There’s also the supply-chain angle: securing key materials like T-
Anthropic · 2 atoms
  • 🟦 Claude Tag 与 Salesforce 的竞争关系 · 竞争格局
    When Anthropic on Tuesday launched a high-profile AI product for businesses that use Slack, some employees at Salesforce, which owns Slack, were confused. In social media posts, Salesforce promoted An
  • 🟥 3 万亿估值下 IPO 可能成流动性抽水 · 估值争议/融资
    SmutCap: Will be interesting if there is any slowdown to arr growth from token :no_entry_sign: maxing, gov intervention. Also unclear if they blast out an IPO this year on current schedule, would imag
CoWoS-R · 1 atoms
  • 🟥 成为 AI CPU 交易新瓶颈 · 技术路线/供给瓶颈
    CoWoS-R is becoming the AI CPU trade. Vera CPU and AMD Venice CPU are why CoWoS-R matters. If agentic AI drives more CPU demand, this becomes another bottleneck people are not fully modeling yet.
GUC · 1 atoms
  • 🟥 CoWoS booking 预期增长 2027 · 产能/出货预测
    MS expects GUC CoWoS booking to rise from 14k wafers in 2026 to 60k wafers in 2027, up more than 4x YoY.
  • 🟥 die 效率与价值差异 · 产品路线/价值比较
    MS assumes Broadcom Sunfish uses around 12 chips per CoWoS wafer, while MediaTek Zebrafish uses around 20 chips per wafer. That is why Broadcom can have more value with similar-ish unit volume. Bigger
MediaTek · 1 atoms
  • 🟥 2027 年 2nm TPU v9 出货预期 · 出货预测/产品路线
    MediaTek could ship around 400k units of 2nm TPU v9 through Intel EMIB-T in 2027.
AVGO · 1 atoms
  • 🟥 ASIC 客户扩张与 Jalapeno 进展 · 客户扩展/产品路线
    CLSA’s custom ASIC stack includes Google, Anthropic, OpenAI, Meta, ByteDance and potentially another undisclosed customer. The market is too focused on Google share loss and not enough on the broader
MU · 1 atoms
  • 🟥 政治化风险压制估值 · 估值/政策风险
    If memory price becomes a politically involved issue, USG starts to pressure memory producers into production expansion sooner, would this press MU multiple but works for AMAT? Also, DRAM ETF is addin
存储器板块 · 1 atoms
  • 🟥 AI 交易双向风险承压 · 宏观风险/估值
    Two things that could kill the AI trade: hyperscalers cutting capex and FOMC embarking on a series of hikes. Memory is the one pushing on both.
  • 🟥 护城河差异:原子级材料科学 vs 设备集成 · 护城河/估值框架
    This new narrative deems unimportant the fact that one group has tackled, over decades of know-how and proprietary IP, atomic level material science into a product that's near impossible and extremely
LTA · 1 atoms
  • 🟥 不是结构性护城河,而是价格封顶 · 估值框架/合约结构
    An LTA narrows the volume distribution at the cost of capping price upside (basically collaring ASP), it holds best when least needed, and degrades precisely when the cycle turns and a new entrant rai
davii · 1 atoms
  • 🟥 LTA 改变了周期性的反驳 · 估值框架
    I agree with ur empirical observation about earnings vol, but that's in past cycles. What is a fair valuation gap then, in a forward-looking sense now that LTAs are real?
Terafab · 1 atoms
  • 🟦 专注制造存储芯片 · 技术路线/产能
    In this week’s All In podcast, it’s mentioned that Terafab has a focus on manufacturing memory chips.
CXMT · 1 atoms
  • 🟥 产能 2025 与未来扩产预期 · 产能预测/替代主线
    CXMT capacity reached 280kwpm in 2025, and domestic DRAM capacity could at least double over the next 3-5 years. HBM equipment intensity is also 30-50% higher than standard DRAM.
YMTC · 1 atoms
  • 🟥 产能 2025 与未来扩产预期 · 产能预测/替代主线
    YMTC capacity reached 160kwpm in 2025, and domestic NAND capacity could at least double. 3D NAND needs a lot of etch and deposition, which account for 55-60% of total capex in a 3D NAND line.
AMEC · 1 atoms
  • 🟥 2026 年订单预期 · 出货预测/替代主线
    CLSA thinks AMEC orders can grow 48% YoY in 2026 and NAURA memory customer orders can grow 50%+ YoY. This is a very direct China AI + memory + localisation trade.
Naura · 1 atoms
  • 🟥 2026 年存储客户订单预期 · 出货预测/替代主线
    CLSA thinks AMEC orders can grow 48% YoY in 2026 and NAURA memory customer orders can grow 50%+ YoY. This is a very direct China AI + memory + localisation trade.
USG · 1 atoms
  • 🟥 可能投资前沿 AI 实验室 · 政策/宏观格局
    Does anyone have any read on possible USG investment in frontier labs? Feels like with recent uncertainties surrounding open source and DC costs that this is only going one way, and that is for an inv
Cloudflare · 1 atoms
  • 🟦 CEO Prince 框架:AI 消灭测量者角色 · 管理层表态/企业结构
    Cloudflare’s engineering headcount surged 45 percent in the weeks after the company cut 1,100 jobs in May... CEO Matthew Prince confirmed the trend and offered a framework for understanding it: every
SUMMARY COVERAGE

Market chatter · 总结覆盖率

历史口径未留档
该日生成时尚未记录覆盖率;未知不记为 0。