← 市场观察

2026-11-04

中期选举之后，AI半导体怎么布局？

选举卖压已经出清，但资本开支回报和加息这两座山还在——用技术、供应链、宏观三个镜头，找出AI半导体里绕不开的公司

一句话结论：选举带来的卖压基本已经出清，但AI股这轮调整的真正原因——资本开支回报的疑虑和美联储加息——并没有因为选举而改变。所以现在适合列清单、分批试探，而不是一把梭。

11月3日中期选举，民主党同时拿下参众两院。周三早盘，AI数据中心相关股票应声下跌，新云（neocloud）跌得最狠。但不到半天，美股就收复失地，抄底资金涌回科技股。

这一跌一弹说明了一件事：市场怕的是不确定性，结果出来，政治风险反而落地了。接下来真正决定AI股方向的，是另外两件事。

一、发生了什么

选举只是导火索，AI篮子早在选前就已经跌了24%。

时间	事件	影响
11月4日	民主党拿下两院；CoreWeave跌约8%，Iren、Nebius各跌约6%	盘中反转，全球股市收复跌幅
10月30日	Intel第四季度指引低于预期，单日跌约28.6%，1974年以来最差	费城半导体指数单日跌7.1%，芯片股市值蒸发约1万亿美元
10月末	四大云厂商财报：Alphabet单季资本开支514亿美元、自由现金流转负；四家2026年开支合计约7850亿美元（穆迪估算）	Alphabet单日跌12.6%，上市以来最差；Meta跌约9%；亚马逊盘后跌超7%
10月全月	高盛AI篮子当月跌15%，较高点回撤24%以上	选举前已进入技术性熊市
9月16日	美联储加息25个基点至3.75%–4.00%，2023年以来首次加息	点阵图暗示年内还有一次

拆开看，有三股力量叠在一起：

• 政治风险：市场担心民主党会收紧数据中心审批、缩减科技税收优惠。但高盛的判断是：可能限制税收优惠，全国性禁令不太可能。这是可定价的风险，不是生存风险。
• 回报疑虑：云厂商并没有削减开支，反而在加。市场惩罚的是“开支涨了、收入没跟上”的公司，奖励的是云收入加速的公司。
• 利率逆风：通胀粘性让美联储重新进入加息通道。对高估值成长股来说，这压住了估值天花板。

选举只解决了第一个问题。后两个，还得看财报和美联储。

二、三个镜头看清局面

技术和供应链两个镜头都还是绿灯，只有市场/宏观这个镜头亮了黄灯。我做判断时三个镜头缺一不可，这次也一样。

技术镜头：需求是真的。第三季度谷歌云收入增长103%，Azure增长48%，AWS增长39%，都在加速。英伟达管理层称下一财年指引仍是“受供给约束”，台积电、美光、Marvell也都说需求跑在供给前面。AI不是没用，而是赚钱速度被市场拿放大镜在看。

供应链镜头：钱还在往上游流。四大云厂商2026年资本开支合计约7850亿美元（穆迪估算），2027年预计逼近1万亿；Alphabet已表示明年开支会“大幅增加”。其中一部分是零部件涨价，今年早些时候微软和Meta都把上调开支部分归因于此。更关键的是：数据中心审批收紧影响的是“建在哪里、电从哪里来”，而芯片可以装进任何地方的机房。越往上游的瓶颈，政策敏感度越低。

市场/宏观镜头：黄灯。美联储处于加息周期，市场押12月再加一次。中期选举年历史上下半年常见一次大回撤，而选后市场通常上涨——但历史规律没有把加息周期这个变量算进去。

结论：基本面没坏，估值压力未解除。这是“分批布局、等确认再加仓”的局面。

三、哪些公司真正绕不开

结论先说：真正没有替代品的只有四个环节——EUV光刻（ASML）、先进制程和CoWoS封装（台积电）、芯片设计软件（Synopsys和Cadence）、ABF绝缘膜（味之素）。英伟达反而不在这一档，因为定制ASIC已经是现实可用的替代品。

“绕不开”不是感觉，我用四条标准来判断：

1. 有几家能做。只有一家是垄断，两三家是寡头；十家以上就谈不上瓶颈。
2. 换掉要多久。就算有竞争对手，客户切换供应商要重新验证。切换周期超过两年，就等于绑定。
3. 现在缺不缺货。交期变长、价格上涨、客户预付定金，说明它有定价权。
4. 瓶颈还能撑多久。产能扩张会缓解缺货。一个明年就松动的瓶颈，和一个十年都绕不开的瓶颈，价值完全不同。

前两条决定“是不是瓶颈”，后两条决定“现在值多少钱”。

瓶颈地图

第一档的四个环节分布在设计、制造、封装三个阶段，而英伟达所在的AI加速器环节属于第三档。通电环节没有第一档，但燃气轮机的交期最长。

逐环节打分

13个环节里，有四个实际上只有一到两家能做，而且几年内没有可用的替代品。表中数据截至2026年11月，标注（日）（韩）（德）（法）的是不在美国或台湾上市的公司。

环节	谁控制	几家能做	换掉要多久	缺货证据	瓶颈还能撑多久
EUV光刻机	ASML	1家，全球唯一商用EUV供应商	没有替代	2025年底积压订单388亿欧元，创纪录	十年以上
先进制程（3nm以下）	台积电	实际上1家，先进3nm产能占比超70%	3–5年，三星和英特尔仍在追良率	AI占台积电3nm产能从5%升到36%	长
CoWoS先进封装	台积电（溢出单给日月光、安靠）	台积电为主	1–2年，外包厂做不了最高端的CoWoS-L	2026年产能售罄，交期52–78周	供需缺口预计年底从20%收窄到10%，2027年缓解
EDA芯片设计软件	Synopsys、Cadence	2家，合计约85%	5年以上；2nm以下离开商用EDA无法设计	不靠缺货，靠订阅和每代制程涨价	十年以上
ABF绝缘膜	味之素（日）	1家，约95%	新供应商认证至少18–24个月	产线满载，考虑提价30%以上	至少到2027年底；玻璃基板是远期威胁
ABF载板	Ibiden（日）、欣兴、新光电气（日）等	约5家，前三家占约四分之三	1–2年	现货价涨超30%，欣兴产能利用率超90%	到2027年前后
HBM高带宽内存	SK海力士（韩）、三星（韩）、美光	3家	1–2年，卡在客户认证	三家2026年全部售罄；美光客户只拿到订单量的60–70%	2027年三星放量后价格可能松动
制程检测与量测	KLA	一家独大，份额约55–58%	3年以上	交期拉长；2026年过程控制业务增长超20%	长，制程越复杂用得越多
磷化铟激光器（光模块核心）	Lumentum、Coherent	2家主导	2年以上，磷化铟晶圆厂扩产慢	Lumentum出货缺口约25–30%；英伟达给两家各投资20亿美元锁产能	有分歧：研究机构称2026年底可缓解，Lumentum称到了2028年仍缺
燃气轮机	GE Vernova、西门子能源（德）、三菱重工（日）	3家	没有快速替代	GEV积压加预留116GW，今天下单约2031年交货	到2030年前后；但部分是重复下单
AI加速器	英伟达；博通、Marvell（定制ASIC）	1家主导，但有ASIC替代	1–3年，谷歌TPU已大规模部署	Blackwell和Rubin供不应求	份额从85%以上向约75%回落
刻蚀与沉积设备	Lam、应用材料、东京电子（日）	3家竞争	1–3年	管理层称供应链约束压低了指引	跟随设备周期
液冷与供电	Vertiv、施耐德（法）等	多家	1年以内	Vertiv三季度订单同比翻倍以上	竞争会加剧

三档结论

第一档：几乎不可替代。只有一到两家能做，竞争者几年内追不上。不管AI开支流向GPU还是ASIC、芯片建在德州还是中东，都要从这里过。

• ASML：每一颗3nm以下的芯片都要用EUV曝光，而EUV只有它一家能造。光学镜头来自蔡司，光源来自通快，整个供应链嵌在它身上，竞争者连门槛都摸不到。存储厂扩产HBM也在买EUV，2025年第四季度新增订单约56%来自存储客户。
• 台积电：一个公司占了两个瓶颈——先进制程和CoWoS封装。英伟达、博通、AMD、谷歌、亚马逊的芯片都在这里流片。谁赢芯片设计之战，台积电都收钱。
• Synopsys、Cadence：最常被忽略的瓶颈。每一颗AI芯片，包括每一颗定制ASIC，都要用它们的软件设计和验证。云厂商自己做芯片越多，它们的客户就越多。利润不靠缺货，而靠切换成本——换工具要重新验证整套设计流程。
• 味之素（日本上市）：做味精起家的公司，控制全球约95%的ABF绝缘膜。没有它的膜，载板做不出来，GPU就没法封装出货。不在美国或台湾上市，但值得知道这个瓶颈在哪。

第二档：寡头且正在缺货。两三家能做，现在供不应求、有定价权。但产能在扩，要盯紧瓶颈什么时候松动。

• KLA：制程检测份额是第二名的六倍多，接近第一档。制程越先进、HBM堆叠越高，需要的检测步骤越多，这是结构性增长。放在第二档，是因为应用材料等公司在部分细分领域仍有竞争力。
• HBM（SK海力士、三星、美光）：三家今年全部售罄。但存储是周期行业，三星HBM4认证后，2027年价格可能松动。美光今年已涨约240%，定价权很可能已经反映在股价里。
• Lumentum、Coherent：1.6T光模块里的激光器只能用磷化铟做——硅本身不能发光。每个1.6T模块用8颗磷化铟激光器，是800G的两倍。另外值得注意：彭博社报道Marvell正在谈收购Lumentum。
• GE Vernova：电是数据中心最硬的约束，全球只有三家能造大型燃气轮机，GEV已经在卖约2031年的交货位置。要注意其116GW里有63GW是付费预留而非确定订单，且客户可能同时向几家下单。
• ABF载板（欣兴等）：在味之素之下一层，接近满产。欣兴是台湾上市（3037.TW），符合研究范围。

第三档：强势但有替代者。是好公司，但不是“绕不开”。

• 英伟达：仍占AI加速器大部分份额，CUDA生态深厚。但谷歌TPU、亚马逊Trainium已经证明大客户可以把部分工作量搬走。彭博行业研究预计到2033年定制ASIC年均增长44.6%，GPU为16.1%。它的护城河另有来源（见下节）。
• 博通：定制ASIC份额约70%，另有Tomahawk交换芯片。但ASIC设计客户可以转向Marvell或联发科，这是和台积电不同的地方。
• Lam：刻蚀和沉积是存储扩产的主力设备，但与应用材料、东京电子三家竞争。走强更多来自设备周期，而不是瓶颈地位。
• Vertiv：液冷是刚需，但供应商众多、技术门槛低于芯片，增长越快越会引来竞争。

反直觉的发现：英伟达的护城河是“买断瓶颈”

英伟达自己不是瓶颈，但它把钱和订单压在了大部分瓶颈上。竞争对手就算设计出一样好的芯片，也很难拿到足够的产能来生产：

• CoWoS：业内估计英伟达锁定了台积电2026年CoWoS产能的约50–60%。
• HBM4：三家产能售罄，其中大头给了英伟达的Rubin。
• 磷化铟激光器：2026年3月给Lumentum和Coherent各投资20亿美元并签长期供货协议。
• 总规模：英伟达的供货承诺达到2790亿美元。

这带来两个启示。第一，只要瓶颈还在，英伟达的领先就比技术本身更难被追上；瓶颈松动之日，反而是ASIC抢份额加速之时。所以盯紧CoWoS和HBM的供需缺口，就是在盯英伟达的护城河。

第二，2790亿美元的承诺是双刃剑。需求好时是护城河，需求转弱时是负担。不管哪种情况，这笔钱的另一端都是第一、二档的公司。

绕不开不等于现在就该买

瓶颈强度和股价表现错位最明显的是Synopsys：第一档的瓶颈，今年股价却是负的。反过来，美光和Lumentum的缺货已经大幅反映在股价里。

整个板块的估值已经压缩：10月30日标普半导体指数远期市盈率降至14.7倍，一年前约为33倍，已低于标普500的18.9倍。但低市盈率的前提是盈利预期不下修。

公司	档位	近期股价	估值参考	错位判断
Synopsys	第一档	今年以来约-3%（10月27日）	未来12个月约28倍，低于Cadence的35.5倍；维权基金Elliott已介入	瓶颈未被定价，整合Ansys是主要分歧
ASML	第一档	近一年约+45%（11月3日）	远期约31倍（9月24日）	垄断溢价已有，但不离谱
台积电	第一档	今年以来约+45%（10月底）	2026年增长预期三次上调至约44%	增长和股价大致匹配
英伟达	第三档（但买断瓶颈）	近一年基本持平（11月3日）	明年盈利的约13.5倍	最便宜的龙头，市场押的是ASIC分流
博通	第三档	随板块回调	明年盈利的约18倍	增长指引强，估值合理
Coherent	第二档	10月30日单日跌11%	远期约22倍，低于五年均值24倍，6月曾达57倍	缺货仍在，估值已回到均值以下
Lumentum	第二档	今年以来超+150%	远期约34倍；有收购传闻	缺货已大幅定价
美光	第二档	今年以来约+240%	周期股，市盈率低不代表便宜	缺货已大幅定价，小心2027年价格拐点

两点提醒：

• 估值数字来自不同来源和日期，只用来比较相对位置，不要当成精确值。
• 用Minervini的框架看，“滞涨”本身也是信号。Synopsys这样的名字相对强度弱，结构上再好不等于股价会动。先放进观察清单，等它自己用走势证明。

本周就有验证：Coherent 11月4日盘后、Lumentum 11月5日盘后发布财报。如果激光器缺口明显收窄，光通信就要从第二档往下调。

四、其他板块怎么配

非AI部分的任务是降低和AI仓位的相关性，而不是再找一个高Beta故事。

板块	方向	理由
医疗	偏多	分析师认为民主党掌握两院时医疗受益最大；选前已是标普最强板块之一，注意已部分计价
电力 / 电网	中性偏多	AI用电需求和电网投资是长期趋势；但审批收紧可能推迟部分需求，受监管公用事业和电网设备比独立发电商稳
国防（尤其无人机）	观望	国会换手后有预算风险；AeroVironment、Kratos已是跌幅最大的一批
金融	观望	资本要求和消费者保护监管可能收紧
现金 / 短债	保留	加息周期里现金有利息，也是等确认信号时的弹药

政策层面的大前提是：分裂政府很难快速通过重大立法。市场今天的反转本身就是在定价“现状不会很快改变”，所以板块轮动不宜做得太激进。

五、怎么布局：纪律比选股更重要

按Minervini的做法，现在是列清单的时候，不是重仓的时候。具体分五步：

1. 不在下跌趋势里抄底。等纳指或费城半导体指数出现放量上涨的确认日（follow-through day），再开始建仓。
2. 找回调中最抗跌的股票。指数跌了20%以上，有些股票离高点只有10%左右、RS线没破位，它们往往就是下一轮领涨股。
3. 分档建仓。先用1/4仓试探，赚钱了再加到半仓、满仓；单笔止损7–8%。
4. 看内部人动向做交叉验证。大跌之后，半导体公司高管有没有在公开市场自己掏钱买入，是很强的佐证信号。
5. 留足现金。加息周期里现金不是浪费，是等确认信号时的弹药。

六、接下来盯什么

两个催化剂决定这轮调整什么时候结束：英伟达下一份财报，和12月FOMC。

• 英伟达财报：验证“开支能不能赚回来”的故事。重点看数据中心收入指引和供给约束的措辞。
• 12月FOMC：决定估值天花板。如果加息后表态转向观望，成长股的压力会减轻不少。
• 云厂商的现金流：开支不减没关系，关键是云收入增速能不能跟上。
• 州级数据中心政策：尤其德州和佐治亚新州长上任后的审批和电价政策。

本文仅为个人研究笔记和思考框架，不构成任何投资建议。文中数据截至2026年11月4日，投资有风险，请独立判断。

资料来源

• AI data center stocks drop as Democrats sweep U.S. midterms – Investing.com
• How Markets Reacted to Democrats’ Win – Bloomberg
• Midterm elections could jolt stocks – CNBC
• Chip stocks lose $1 trillion as Intel sinks – Benzinga
• Intel’s worst drop since 1974 – Value Add VC
• VIX whipsaw on sticky inflation and AI capex – Yahoo Finance
• Alphabet sheds $500 billion in a single day despite cloud growth of 103% – Motley Fool
• Alphabet Q3 2026: worst day on record explained – Pomegra
• Big Tech earnings: Azure +48%, Alphabet -12.6% – Shattered
• Azure hits 48%, hyperscaler capex outlook – Tech Times
• Alphabet, Microsoft, Meta earnings will judge AI capex – AI Weekly
• Fed hikes, hints at more – InvestingLive
• Why health care stocks are rallying – Axios
• AI policy fears grip traders before midterms – Advisor Perspectives
• CoWoS supply-demand gap seen narrowing – TrendForce
• TSMC's CoWoS bottleneck explained – INDmoney
• TSMC CoWoS packaging bottleneck – Live in the Future
• AI demand keeps 3nm and CoWoS tight through 2027 – Pulse
• TSMC to quadruple advanced packaging capacity – TokenRing
• ASML projects $71B revenue by 2030 – Tom's Hardware
• ASML Q4 bookings double – Rallies
• Nasdaq-100 Inside the Index: ASML – Nasdaq
• EDA Market Primer – SemiAnalysis
• Synopsys: how the business works – Belcourt
• The state of ABF substrates in 2026 – Tom's Hardware
• The ABF substrate bottleneck – Wing VC
• Ajinomoto to raise ABF film prices – OnMSFT
• Substrate shortages and Unimicron expansion – T.Media
• Micron HBM storyline – Silicon Analysts
• KLA: how the business works – Belcourt
• Can process control intensity drive KLA's growth – Nasdaq
• InP shortage emerges as AI optical bottleneck – TrendForce
• The indium phosphide squeeze – Abhay Terminal
• Lumentum touts AI optics boom – MarketBeat
• Lumentum InP AI optics shortage – XenoSpectrum
• The gas turbine shortage just became AI's biggest constraint – OilPrice
• US power firms scramble for transformers and turbines – The Next Web
• NVIDIA vs Broadcom – Metavert
• If AI spending slows in 2027 – 24/7 Wall St.
• Lam Research FY27 Q1 analyst views – StockHub
• Vertiv Q3 2026 earnings call – The Singju Post
• Semiconductor stocks are cheaper than the S&P 500 – Benzinga
• My top 3 AI stock picks for November – Motley Fool
• Semiconductor stocks theme page – StockTitan
• Why I keep buying Taiwan Semiconductor – Yahoo Finance
• What just happened to Synopsys stock – Trefis
• Synopsys valuation and news – Parqet
• Coherent and Lumentum initiated at Buy – TIKR
• Coherent stock fell 11% – TIKR
• Lumentum quote and news – StockAnalysis
• Lumentum forward P/E – StockAnalysis

← 市場觀察

2026-11-04

中期選舉之後，AI半導體怎麼佈局？

選舉賣壓已經出清，但資本開支回報和升息這兩座山還在——用技術、供應鏈、宏觀三個鏡頭，找出AI半導體裡繞不開的公司

一句話結論：選舉帶來的賣壓基本已經出清，但AI股這輪調整的真正原因——資本開支回報的疑慮和聯準會升息——並沒有因為選舉而改變。所以現在適合列清單、分批試探，而不是一把梭。

11月3日中期選舉，民主黨同時拿下參眾兩院。週三早盤，AI資料中心相關股票應聲下跌，新雲（neocloud）跌得最狠。但不到半天，美股就收復失地，抄底資金湧回科技股。

這一跌一彈說明了一件事：市場怕的是不確定性，結果出來，政治風險反而落地了。接下來真正決定AI股方向的，是另外兩件事。

一、發生了什麼

選舉只是導火索，AI籃子早在選前就已經跌了24%。

時間	事件	影響
11月4日	民主黨拿下兩院；CoreWeave跌約8%，Iren、Nebius各跌約6%	盤中反轉，全球股市收復跌幅
10月30日	Intel第四季度指引低於預期，單日跌約28.6%，1974年以來最差	費城半導體指數單日跌7.1%，晶片股市值蒸發約1萬億美元
10月末	四大雲廠商財報：Alphabet單季資本開支514億美元、自由現金流轉負；四家2026年開支合計約7850億美元（穆迪估算）	Alphabet單日跌12.6%，上市以來最差；Meta跌約9%；亞馬遜盤後跌超7%
10月全月	高盛AI籃子當月跌15%，較高點回撤24%以上	選舉前已進入技術性熊市
9月16日	聯準會升息25個基點至3.75%–4.00%，2023年以來首次升息	點陣圖暗示年內還有一次

拆開看，有三股力量疊在一起：

• 政治風險：市場擔心民主黨會收緊資料中心審批、縮減科技稅收優惠。但高盛的判斷是：可能限制稅收優惠，全國性禁令不太可能。這是可定價的風險，不是生存風險。
• 回報疑慮：雲廠商並沒有削減開支，反而在加。市場懲罰的是“開支漲了、收入沒跟上”的公司，獎勵的是雲收入加速的公司。
• 利率逆風：通膨粘性讓聯準會重新進入升息通道。對高估值成長股來說，這壓住了估值天花板。

選舉只解決了第一個問題。後兩個，還得看財報和聯準會。

二、三個鏡頭看清局面

技術和供應鏈兩個鏡頭都還是綠燈，只有市場/宏觀這個鏡頭亮了黃燈。我做判斷時三個鏡頭缺一不可，這次也一樣。

技術鏡頭：需求是真的。第三季度谷歌雲收入增長103%，Azure增長48%，AWS增長39%，都在加速。輝達管理層稱下一財年指引仍是“受供給約束”，台積電、美光、Marvell也都說需求跑在供給前面。AI不是沒用，而是賺錢速度被市場拿放大鏡在看。

供應鏈鏡頭：錢還在往上游流。四大雲廠商2026年資本開支合計約7850億美元（穆迪估算），2027年預計逼近1萬億；Alphabet已表示明年開支會“大幅增加”。其中一部分是零部件漲價，今年早些時候微軟和Meta都把上調開支部分歸因於此。更關鍵的是：資料中心審批收緊影響的是“建在哪裡、電從哪裡來”，而晶片可以裝進任何地方的機房。越往上游的瓶頸，政策敏感度越低。

市場/宏觀鏡頭：黃燈。聯準會處於升息週期，市場押12月再加一次。中期選舉年歷史上下半年常見一次大回撤，而選後市場通常上漲——但歷史規律沒有把升息週期這個變數算進去。

結論：基本面沒壞，估值壓力未解除。這是“分批佈局、等確認再加倉”的局面。

三、哪些公司真正繞不開

結論先說：真正沒有替代品的只有四個環節——EUV光刻（ASML）、先進製程和CoWoS封裝（台積電）、晶片設計軟體（Synopsys和Cadence）、ABF絕緣膜（味之素）。輝達反而不在這一檔，因為定製ASIC已經是現實可用的替代品。

“繞不開”不是感覺，我用四條標準來判斷：

1. 有幾家能做。只有一家是壟斷，兩三家是寡頭；十家以上就談不上瓶頸。
2. 換掉要多久。就算有競爭對手，客戶切換供應商要重新驗證。切換週期超過兩年，就等於繫結。
3. 現在缺不缺貨。交期變長、價格上漲、客戶預付定金，說明它有定價權。
4. 瓶頸還能撐多久。產能擴張會緩解缺貨。一個明年就鬆動的瓶頸，和一個十年都繞不開的瓶頸，價值完全不同。

前兩條決定“是不是瓶頸”，後兩條決定“現在值多少錢”。

瓶頸地圖

第一檔的四個環節分佈在設計、製造、封裝三個階段，而輝達所在的AI加速器環節屬於第三檔。通電環節沒有第一檔，但燃氣輪機的交期最長。

逐環節打分

13個環節裡，有四個實際上只有一到兩家能做，而且幾年內沒有可用的替代品。表中資料截至2026年11月，標註（日）（韓）（德）（法）的是不在美國或台灣上市的公司。

環節	誰控制	幾家能做	換掉要多久	缺貨證據	瓶頸還能撐多久
EUV光刻機	ASML	1家，全球唯一商用EUV供應商	沒有替代	2025年底積壓訂單388億歐元，創紀錄	十年以上
先進製程（3nm以下）	台積電	實際上1家，先進3nm產能佔比超70%	3–5年，三星和英特爾仍在追良率	AI佔台積電3nm產能從5%升到36%	長
CoWoS先進封裝	台積電（溢位單給日月光、安靠）	台積電為主	1–2年，外包廠做不了最高階的CoWoS-L	2026年產能售罄，交期52–78周	供需缺口預計年底從20%收窄到10%，2027年緩解
EDA晶片設計軟體	Synopsys、Cadence	2家，合計約85%	5年以上；2nm以下離開商用EDA無法設計	不靠缺貨，靠訂閱和每代製程漲價	十年以上
ABF絕緣膜	味之素（日）	1家，約95%	新供應商認證至少18–24個月	產線滿載，考慮提價30%以上	至少到2027年底；玻璃基板是遠期威脅
ABF載板	Ibiden（日）、欣興、新光電氣（日）等	約5家，前三家佔約四分之三	1–2年	現貨價漲超30%，欣興產能利用率超90%	到2027年前後
HBM高頻寬記憶體	SK海力士（韓）、三星（韓）、美光	3家	1–2年，卡在客戶認證	三家2026年全部售罄；美光客戶只拿到訂單量的60–70%	2027年三星放量後價格可能鬆動
製程檢測與量測	KLA	一家獨大，份額約55–58%	3年以上	交期拉長；2026年過程控制業務增長超20%	長，製程越複雜用得越多
磷化銦鐳射器（光模組核心）	Lumentum、Coherent	2家主導	2年以上，磷化銦晶圓廠擴產慢	Lumentum出貨缺口約25–30%；輝達給兩家各投資20億美元鎖產能	有分歧：研究機構稱2026年底可緩解，Lumentum稱到了2028年仍缺
燃氣輪機	GE Vernova、西門子能源（德）、三菱重工（日）	3家	沒有快速替代	GEV積壓加預留116GW，今天下單約2031年交貨	到2030年前後；但部分是重複下單
AI加速器	輝達；博通、Marvell（定製ASIC）	1家主導，但有ASIC替代	1–3年，谷歌TPU已大規模部署	Blackwell和Rubin供不應求	份額從85%以上向約75%回落
刻蝕與沉積設備	Lam、應用材料、東京電子（日）	3家競爭	1–3年	管理層稱供應鏈約束壓低了指引	跟隨設備週期
液冷與供電	Vertiv、施耐德（法）等	多家	1年以內	Vertiv三季度訂單同比翻倍以上	競爭會加劇

三檔結論

第一檔：幾乎不可替代。只有一到兩家能做，競爭者幾年內追不上。不管AI開支流向GPU還是ASIC、晶片建在德州還是中東，都要從這裡過。

• ASML：每一顆3nm以下的晶片都要用EUV曝光，而EUV只有它一家能造。光學鏡頭來自蔡司，光源來自通快，整個供應鏈嵌在它身上，競爭者連門檻都摸不到。儲存廠擴產HBM也在買EUV，2025年第四季度新增訂單約56%來自儲存客戶。
• 台積電：一個公司佔了兩個瓶頸——先進製程和CoWoS封裝。輝達、博通、AMD、谷歌、亞馬遜的晶片都在這裡流片。誰贏晶片設計之戰，台積電都收錢。
• Synopsys、Cadence：最常被忽略的瓶頸。每一顆AI晶片，包括每一顆定製ASIC，都要用它們的軟體設計和驗證。雲廠商自己做晶片越多，它們的客戶就越多。利潤不靠缺貨，而靠切換成本——換工具要重新驗證整套設計流程。
• 味之素（日本上市）：做味精起家的公司，控制全球約95%的ABF絕緣膜。沒有它的膜，載板做不出來，GPU就沒法封裝出貨。不在美國或台灣上市，但值得知道這個瓶頸在哪。

第二檔：寡頭且正在缺貨。兩三家能做，現在供不應求、有定價權。但產能在擴，要盯緊瓶頸什麼時候鬆動。

• KLA：製程檢測份額是第二名的六倍多，接近第一檔。製程越先進、HBM堆疊越高，需要的檢測步驟越多，這是結構性增長。放在第二檔，是因為應用材料等公司在部分細分領域仍有競爭力。
• HBM（SK海力士、三星、美光）：三家今年全部售罄。但儲存是週期行業，三星HBM4認證後，2027年價格可能鬆動。美光今年已漲約240%，定價權很可能已經反映在股價裡。
• Lumentum、Coherent：1.6T光模組裡的鐳射器只能用磷化銦做——矽本身不能發光。每個1.6T模組用8顆磷化銦鐳射器，是800G的兩倍。另外值得注意：彭博社報道Marvell正在談收購Lumentum。
• GE Vernova：電是資料中心最硬的約束，全球只有三家能造大型燃氣輪機，GEV已經在賣約2031年的交貨位置。要注意其116GW裡有63GW是付費預留而非確定訂單，且客戶可能同時向幾家下單。
• ABF載板（欣興等）：在味之素之下一層，接近滿產。欣興是台灣上市（3037.TW），符合研究範圍。

第三檔：強勢但有替代者。是好公司，但不是“繞不開”。

• 輝達：仍佔AI加速器大部分份額，CUDA生態深厚。但谷歌TPU、亞馬遜Trainium已經證明大客戶可以把部分工作量搬走。彭博行業研究預計到2033年定製ASIC年均增長44.6%，GPU為16.1%。它的護城河另有來源（見下節）。
• 博通：定製ASIC份額約70%，另有Tomahawk交換晶片。但ASIC設計客戶可以轉向Marvell或聯發科，這是和台積電不同的地方。
• Lam：刻蝕和沉積是儲存擴產的主力設備，但與應用材料、東京電子三家競爭。走強更多來自設備週期，而不是瓶頸地位。
• Vertiv：液冷是剛需，但供應商眾多、技術門檻低於晶片，增長越快越會引來競爭。

反直覺的發現：輝達的護城河是“買斷瓶頸”

輝達自己不是瓶頸，但它把錢和訂單壓在了大部分瓶頸上。競爭對手就算設計出一樣好的晶片，也很難拿到足夠的產能來生產：

• CoWoS：業內估計輝達鎖定了台積電2026年CoWoS產能的約50–60%。
• HBM4：三家產能售罄，其中大頭給了輝達的Rubin。
• 磷化銦鐳射器：2026年3月給Lumentum和Coherent各投資20億美元並籤長期供貨協議。
• 總規模：輝達的供貨承諾達到2790億美元。

這帶來兩個啟示。第一，只要瓶頸還在，輝達的領先就比技術本身更難被追上；瓶頸鬆動之日，反而是ASIC搶份額加速之時。所以盯緊CoWoS和HBM的供需缺口，就是在盯輝達的護城河。

第二，2790億美元的承諾是雙刃劍。需求好時是護城河，需求轉弱時是負擔。不管哪種情況，這筆錢的另一端都是第一、二檔的公司。

繞不開不等於現在就該買

瓶頸強度和股價表現錯位最明顯的是Synopsys：第一檔的瓶頸，今年股價卻是負的。反過來，美光和Lumentum的缺貨已經大幅反映在股價裡。

整個板塊的估值已經壓縮：10月30日標普半導體指數遠期市盈率降至14.7倍，一年前約為33倍，已低於標普500的18.9倍。但低市盈率的前提是盈利預期不下修。

公司	檔位	近期股價	估值參考	錯位判斷
Synopsys	第一檔	今年以來約-3%（10月27日）	未來12個月約28倍，低於Cadence的35.5倍；維權基金Elliott已介入	瓶頸未被定價，整合Ansys是主要分歧
ASML	第一檔	近一年約+45%（11月3日）	遠期約31倍（9月24日）	壟斷溢價已有，但不離譜
台積電	第一檔	今年以來約+45%（10月底）	2026年增長預期三次上調至約44%	增長和股價大致匹配
輝達	第三檔（但買斷瓶頸）	近一年基本持平（11月3日）	明年盈利的約13.5倍	最便宜的龍頭，市場押的是ASIC分流
博通	第三檔	隨板塊回撥	明年盈利的約18倍	增長指引強，估值合理
Coherent	第二檔	10月30日單日跌11%	遠期約22倍，低於五年均值24倍，6月曾達57倍	缺貨仍在，估值已回到均值以下
Lumentum	第二檔	今年以來超+150%	遠期約34倍；有收購傳聞	缺貨已大幅定價
美光	第二檔	今年以來約+240%	週期股，市盈率低不代表便宜	缺貨已大幅定價，小心2027年價格拐點

兩點提醒：

• 估值數字來自不同來源和日期，只用來比較相對位置，不要當成精確值。
• 用Minervini的框架看，“滯漲”本身也是訊號。Synopsys這樣的名字相對強度弱，結構上再好不等於股價會動。先放進觀察清單，等它自己用走勢證明。

本週就有驗證：Coherent 11月4日盤後、Lumentum 11月5日盤後發布財報。如果鐳射器缺口明顯收窄，光通訊就要從第二檔往下調。

四、其他板塊怎麼配

非AI部分的任務是降低和AI倉位的相關性，而不是再找一個高Beta故事。

板塊	方向	理由
醫療	偏多	分析師認為民主黨掌握兩院時醫療受益最大；選前已是標普最強板塊之一，注意已部分計價
電力 / 電網	中性偏多	AI用電需求和電網投資是長期趨勢；但審批收緊可能推遲部分需求，受監管公用事業和電網設備比獨立發電商穩
國防（尤其無人機）	觀望	國會換手後有預算風險；AeroVironment、Kratos已是跌幅最大的一批
金融	觀望	資本要求和消費者保護監管可能收緊
現金 / 短債	保留	升息週期裡現金有利息，也是等確認訊號時的彈藥

政策層面的大前提是：分裂政府很難快速透過重大立法。市場今天的反轉本身就是在定價“現狀不會很快改變”，所以板塊輪動不宜做得太激進。

五、怎麼佈局：紀律比選股更重要

按Minervini的做法，現在是列清單的時候，不是重倉的時候。具體分五步：

1. 不在下跌趨勢裡抄底。等納指或費城半導體指數出現放量上漲的確認日（follow-through day），再開始建倉。
2. 找回撥中最抗跌的股票。指數跌了20%以上，有些股票離高點只有10%左右、RS線沒破位，它們往往就是下一輪領漲股。
3. 分檔建倉。先用1/4倉試探，賺錢了再加到半倉、滿倉；單筆止損7–8%。
4. 看內部人動向做交叉驗證。大跌之後，半導體公司高管有沒有在公開市場自己掏錢買入，是很強的佐證訊號。
5. 留足現金。升息週期裡現金不是浪費，是等確認訊號時的彈藥。

六、接下來盯什麼

兩個催化劑決定這輪調整什麼時候結束：輝達下一份財報，和12月FOMC。

• 輝達財報：驗證“開支能不能賺回來”的故事。重點看資料中心收入指引和供給約束的措辭。
• 12月FOMC：決定估值天花板。如果升息後表態轉向觀望，成長股的壓力會減輕不少。
• 雲廠商的現金流：開支不減沒關係，關鍵是雲收入增速能不能跟上。
• 州級資料中心政策：尤其德州和佐治亞新州長上任後的審批和電價政策。

本文僅為個人研究筆記和思考框架，不構成任何投資建議。文中資料截至2026年11月4日，投資有風險，請獨立判斷。

資料來源

• AI data center stocks drop as Democrats sweep U.S. midterms – Investing.com
• How Markets Reacted to Democrats’ Win – Bloomberg
• Midterm elections could jolt stocks – CNBC
• Chip stocks lose $1 trillion as Intel sinks – Benzinga
• Intel’s worst drop since 1974 – Value Add VC
• VIX whipsaw on sticky inflation and AI capex – Yahoo Finance
• Alphabet sheds $500 billion in a single day despite cloud growth of 103% – Motley Fool
• Alphabet Q3 2026: worst day on record explained – Pomegra
• Big Tech earnings: Azure +48%, Alphabet -12.6% – Shattered
• Azure hits 48%, hyperscaler capex outlook – Tech Times
• Alphabet, Microsoft, Meta earnings will judge AI capex – AI Weekly
• Fed hikes, hints at more – InvestingLive
• Why health care stocks are rallying – Axios
• AI policy fears grip traders before midterms – Advisor Perspectives
• CoWoS supply-demand gap seen narrowing – TrendForce
• TSMC's CoWoS bottleneck explained – INDmoney
• TSMC CoWoS packaging bottleneck – Live in the Future
• AI demand keeps 3nm and CoWoS tight through 2027 – Pulse
• TSMC to quadruple advanced packaging capacity – TokenRing
• ASML projects $71B revenue by 2030 – Tom's Hardware
• ASML Q4 bookings double – Rallies
• Nasdaq-100 Inside the Index: ASML – Nasdaq
• EDA Market Primer – SemiAnalysis
• Synopsys: how the business works – Belcourt
• The state of ABF substrates in 2026 – Tom's Hardware
• The ABF substrate bottleneck – Wing VC
• Ajinomoto to raise ABF film prices – OnMSFT
• Substrate shortages and Unimicron expansion – T.Media
• Micron HBM storyline – Silicon Analysts
• KLA: how the business works – Belcourt
• Can process control intensity drive KLA's growth – Nasdaq
• InP shortage emerges as AI optical bottleneck – TrendForce
• The indium phosphide squeeze – Abhay Terminal
• Lumentum touts AI optics boom – MarketBeat
• Lumentum InP AI optics shortage – XenoSpectrum
• The gas turbine shortage just became AI's biggest constraint – OilPrice
• US power firms scramble for transformers and turbines – The Next Web
• NVIDIA vs Broadcom – Metavert
• If AI spending slows in 2027 – 24/7 Wall St.
• Lam Research FY27 Q1 analyst views – StockHub
• Vertiv Q3 2026 earnings call – The Singju Post
• Semiconductor stocks are cheaper than the S&P 500 – Benzinga
• My top 3 AI stock picks for November – Motley Fool
• Semiconductor stocks theme page – StockTitan
• Why I keep buying Taiwan Semiconductor – Yahoo Finance
• What just happened to Synopsys stock – Trefis
• Synopsys valuation and news – Parqet
• Coherent and Lumentum initiated at Buy – TIKR
• Coherent stock fell 11% – TIKR
• Lumentum quote and news – StockAnalysis
• Lumentum forward P/E – StockAnalysis

← Market watch

2026-11-04

After the midterms, how to position in AI semiconductors

The election selling is done, but two weights remain — doubts about AI capex returns and Fed hikes. Three lenses (technology, supply chain, macro) to find the AI chip companies the industry cannot route around.

In one sentence: the selling triggered by the election has largely run its course, but the real reasons for this AI pullback — doubts about returns on capital spending and the Fed's rate hikes — have not changed. This is a time to build a watchlist and test the water in small pieces, not to go all in.

In the November 3 midterms, Democrats won both the House and the Senate. AI data-center stocks fell early Wednesday, with neoclouds hit hardest. But within half a day US stocks had recovered their losses as dip buyers came back to tech.

That drop and rebound says one thing: markets fear uncertainty, and once the result was in, the political risk was priced. What decides the direction of AI stocks from here are two other things.

1. What happened

The election was only the trigger. The AI basket had already fallen 24% before the vote.

Date	Event	Effect
Nov 4	Democrats win both chambers; CoreWeave down about 8%, Iren and Nebius about 6% each	Reversed intraday; global stocks recovered their losses
Oct 30	Intel's Q4 guidance misses; stock falls about 28.6%, its worst day since 1974	Philadelphia Semiconductor Index down 7.1%; about $1 trillion of chip market value lost
Late Oct	Big four cloud providers report: Alphabet spends $51.4 billion in the quarter and free cash flow turns negative; the four's 2026 capex totals about $785 billion (Moody's estimate)	Alphabet falls 12.6%, its worst day as a public company; Meta about 9%; Amazon over 7% after hours
October	Goldman Sachs AI basket down 15% in the month, more than 24% below its peak	Already in a bear market before the election
Sep 16	Fed raises rates 25 bp to 3.75%–4.00%, its first hike since 2023	Dot plot points to one more hike this year

Three forces are stacked on top of each other:

• Political risk: investors worry Democrats will tighten data-center permitting and trim tech tax incentives. Goldman's view is that tax incentives may be limited, but a national moratorium is unlikely. That is a risk that can be priced, not an existential one.
• Return doubts: cloud providers are not cutting spending; they are raising it. The market punishes companies whose spending rises faster than revenue, and rewards those whose cloud revenue is accelerating.
• Rate headwind: sticky inflation has put the Fed back on a hiking path. For richly valued growth stocks, that caps how high valuations can go.

The election settled only the first. The other two depend on earnings and the Fed.

2. Three lenses on the picture

Technology and supply chain are still green; only the market/macro lens is yellow. I need all three lenses to agree before acting, and this time is no different.

Technology: the demand is real. In the third quarter Google Cloud revenue grew 103%, Azure 48% and AWS 39%, all accelerating. Nvidia's management says next fiscal year's outlook is still "supply-constrained," and TSMC, Micron and Marvell all say demand is running ahead of supply. AI is not useless; the market is putting the pace of payback under a magnifying glass.

Supply chain: money is still flowing upstream. The big four's 2026 capital spending totals about $785 billion (Moody's estimate) and is expected to approach $1 trillion in 2027; Alphabet says next year's spending will "increase significantly." Part of it is higher component prices: earlier this year Microsoft and Meta both attributed some of their capex increases to this. More importantly, tighter permitting affects where data centers are built and where their power comes from, while chips can go into a data center anywhere. The further upstream the bottleneck, the less sensitive it is to policy.

Market/macro: yellow. The Fed is in a hiking cycle and the market expects another hike in December. Midterm years have historically seen a large drawdown in the second half, and markets usually rise after the vote — but those patterns did not include a hiking cycle.

Bottom line: fundamentals are intact, valuation pressure is not resolved. This calls for building positions in stages and adding only after confirmation.

3. Which companies the industry truly cannot route around

The answer first: only four links have no substitute — EUV lithography (ASML), leading-edge manufacturing and CoWoS packaging (TSMC), chip-design software (Synopsys and Cadence), and ABF insulating film (Ajinomoto). Nvidia is not in this tier, because custom ASICs are already a working substitute.

"Cannot route around" is not a feeling. I use four tests:

1. How many companies can do it. One is a monopoly, two or three an oligopoly; ten or more is not a bottleneck.
2. How long it takes to switch. Even with competitors, customers must re-qualify a new supplier. If switching takes more than two years, the customer is effectively locked in.
3. Is it short today. Longer lead times, rising prices and customer prepayments show pricing power.
4. How long the bottleneck lasts. New capacity eases shortages. A bottleneck that loosens next year is worth far less than one that holds for a decade.

The first two decide whether it is a bottleneck; the last two decide what it is worth now.

The bottleneck map

The four first-tier links sit in design, manufacturing and packaging, while the AI accelerator link, where Nvidia sits, is third tier. The power stage has no first-tier link, but gas turbines have the longest lead times of all.

Scoring each link

Of 13 links, four can in practice be done by only one or two companies, with no usable substitute for years. Data as of November 2026; (JP), (KR), (DE) and (FR) mark companies not listed in the US or Taiwan.

Link	Who controls it	How many can do it	Time to switch	Evidence of shortage	How long it lasts
EUV lithography	ASML	1, the only commercial EUV supplier	No substitute	Record order backlog of €38.8 billion at end-2025	A decade or more
Leading-edge foundry (3nm and below)	TSMC	Effectively 1; over 70% of advanced 3nm capacity	3–5 years; Samsung and Intel still chasing yields	AI's share of TSMC's 3nm capacity rose from 5% to 36%	Long
CoWoS advanced packaging	TSMC (overflow to ASE, Amkor)	Mainly TSMC	1–2 years; outsourcers cannot do top-end CoWoS-L	2026 capacity sold out; lead times 52–78 weeks	Supply gap expected to narrow from 20% to 10% by year-end; easing in 2027
EDA chip-design software	Synopsys, Cadence	2, about 85% combined	5+ years; chips at 2nm and below cannot be designed without commercial EDA	Not about shortages: subscriptions and price rises with each node	A decade or more
ABF insulating film	Ajinomoto (JP)	1, about 95%	At least 18–24 months to qualify a new supplier	Lines fully loaded; price rise of 30%+ under consideration	Through at least end-2027; glass substrates are a long-range threat
ABF substrates	Ibiden (JP), Unimicron, Shinko (JP) and others	About 5; top three hold about three-quarters	1–2 years	Spot prices up over 30%; Unimicron utilization above 90%	Around 2027
HBM memory	SK hynix (KR), Samsung (KR), Micron	3	1–2 years, gated by customer qualification	All three sold out for 2026; Micron customers receive only 60–70% of ordered volume	Prices may loosen in 2027 as Samsung ramps
Process control (inspection and metrology)	KLA	One dominant player, about 55–58% share	3+ years	Lead times lengthening; process control expected to grow 20%+ in 2026	Long; more complex processes need more inspection
Indium phosphide lasers (heart of optical modules)	Lumentum, Coherent	2 dominant	2+ years; InP fabs are slow to expand	Lumentum under-shipping demand by about 25–30%; Nvidia invested $2 billion in each to secure capacity	Disputed: a research firm says it could ease by end-2026, Lumentum says shortages persist into 2028
Gas turbines	GE Vernova, Siemens Energy (DE), Mitsubishi Heavy (JP)	3	No quick substitute	GEV backlog plus reservations at 116 GW; orders placed today deliver around 2031	Around 2030; but some orders are duplicates
AI accelerators	Nvidia; Broadcom, Marvell (custom ASICs)	1 dominant, but ASICs substitute	1–3 years; Google's TPUs are deployed at scale	Blackwell and Rubin demand exceeds supply	Share drifting from 85%+ toward about 75%
Etch and deposition	Lam, Applied Materials, Tokyo Electron (JP)	3 competing	1–3 years	Management says supply constraints held guidance down	Follows the equipment cycle
Liquid cooling and power	Vertiv, Schneider (FR) and others	Many	Under 1 year	Vertiv's Q3 orders more than doubled year on year	Competition will intensify

Three tiers

Tier 1: nearly irreplaceable. Only one or two companies can do it, and competitors cannot catch up for years. Whether AI spending flows to GPUs or ASICs, whether chips are installed in Texas or the Middle East, it passes through here.

• ASML: every chip at 3nm and below needs EUV exposure, and only ASML can build EUV tools. Its optics come from Zeiss and its light source from Trumpf; the whole supply chain is built around it, and competitors cannot get near the threshold. Memory makers expanding HBM are buying EUV too: about 56% of new orders in Q4 2025 came from memory customers.
• TSMC: one company holding two bottlenecks — leading-edge manufacturing and CoWoS packaging. Chips from Nvidia, Broadcom, AMD, Google and Amazon are all made here. Whoever wins the chip-design race, TSMC gets paid.
• Synopsys, Cadence: the most overlooked bottleneck. Every AI chip, including every custom ASIC, is designed and verified with their software. The more chips cloud providers design themselves, the more customers these two have. Their profits come not from shortages but from switching costs: changing tools means re-qualifying an entire design flow.
• Ajinomoto (listed in Japan): a company that started out making MSG controls about 95% of ABF insulating film. Without its film, substrates cannot be made and GPUs cannot be packaged and shipped. It is not listed in the US or Taiwan, but it is worth knowing where this bottleneck sits.

Tier 2: oligopolies in shortage. Two or three companies can do it; supply is short and they have pricing power. But capacity is growing, so watch for when the bottleneck loosens.

• KLA: its process-control share is more than six times the runner-up's, close to tier 1. The more advanced the process and the taller the HBM stack, the more inspection steps are needed — structural growth. It sits in tier 2 because Applied Materials and others still compete in some niches.
• HBM (SK hynix, Samsung, Micron): all three are sold out this year. But memory is cyclical, and once Samsung's HBM4 is qualified, 2027 prices may loosen. Micron is up about 240% this year; its pricing power is likely already in the stock.
• Lumentum, Coherent: the lasers in 1.6T optical modules can only be made from indium phosphide — silicon cannot emit light. Each 1.6T module uses 8 InP lasers, twice as many as 800G. Also worth noting: Bloomberg reports Marvell is in talks to buy Lumentum.
• GE Vernova: power is the hardest constraint on data centers. Only three companies in the world build large gas turbines, and GEV is already selling delivery slots around 2031. Note that 63 GW of its 116 GW are paid reservations rather than firm orders, and customers may order from several makers at once.
• ABF substrates (Unimicron and others): one layer below Ajinomoto, running near full capacity. Unimicron is listed in Taiwan (3037.TW), within the research universe.

Tier 3: strong, but substitutes exist. Good companies, but not ones the industry cannot route around.

• Nvidia: still holds most of the AI accelerator market, with a deep CUDA ecosystem. But Google's TPU and Amazon's Trainium have shown big customers can move part of their workloads. Bloomberg Intelligence expects custom ASICs to grow 44.6% a year through 2033, versus 16.1% for GPUs. Its moat comes from somewhere else (next section).
• Broadcom: about 70% of custom ASICs, plus Tomahawk switch chips. But ASIC customers can turn to Marvell or MediaTek, which is the difference from TSMC.
• Lam: etch and deposition are the workhorse tools of memory expansion, but Lam competes with Applied Materials and Tokyo Electron. Its strength owes more to the equipment cycle than to a bottleneck position.
• Vertiv: liquid cooling is a must-have, but suppliers are many and the technical bar is lower than for chips; the faster it grows, the more competition it draws.

A counterintuitive finding: Nvidia's moat is buying up the bottlenecks

Nvidia is not a bottleneck itself, but it has put its money and orders on most of the bottlenecks. Even a competitor with an equally good chip will struggle to get enough capacity to build it:

• CoWoS: the industry estimates Nvidia has locked up about 50–60% of TSMC's 2026 CoWoS capacity.
• HBM4: all three makers are sold out, with the largest share going to Nvidia's Rubin.
• InP lasers: in March 2026 Nvidia invested $2 billion each in Lumentum and Coherent, with long-term supply agreements.
• Overall: Nvidia's supply commitments total $279 billion.

Two implications. First, as long as the bottlenecks hold, Nvidia's lead is harder to close than the technology alone would suggest; the day they loosen is when ASICs gain share fastest. Watching the CoWoS and HBM supply gaps is watching Nvidia's moat.

Second, the $279 billion of commitments cuts both ways: a moat when demand is strong, a burden when it weakens. Either way, the other end of that money is the tier 1 and tier 2 companies.

Irreplaceable does not mean buy now

The clearest mismatch between bottleneck strength and share price is Synopsys: a tier 1 bottleneck whose stock is down this year. Conversely, Micron's and Lumentum's shortages are already largely in their share prices.

Valuations across the sector have compressed: on October 30 the S&P semiconductor index's forward P/E fell to 14.7x from about 33x a year earlier, below the S&P 500's 18.9x. But a low P/E holds only if earnings estimates are not cut.

Company	Tier	Recent share price	Valuation reference	Mismatch
Synopsys	Tier 1	About -3% year to date (Oct 27)	About 28x next-12-month earnings vs Cadence's 35.5x; activist Elliott involved	Bottleneck not priced; Ansys integration is the main debate
ASML	Tier 1	About +45% over one year (Nov 3)	About 31x forward (Sep 24)	Monopoly premium already there, but not extreme
TSMC	Tier 1	About +45% year to date (late Oct)	2026 growth outlook raised three times to about 44%	Growth and price roughly matched
Nvidia	Tier 3 (but buys the bottlenecks)	Roughly flat over one year (Nov 3)	About 13.5x next year's earnings	Cheapest leader; the market is pricing ASIC share loss
Broadcom	Tier 3	Pulled back with the sector	About 18x next year's earnings	Strong growth outlook, reasonable valuation
Coherent	Tier 2	Down 11% on Oct 30	About 22x forward, below its 5-year average of 24x; 57x in June	Shortage persists; valuation back below average
Lumentum	Tier 2	Up over 150% year to date	About 34x forward; takeover reports	Shortage largely priced
Micron	Tier 2	About +240% year to date	Cyclical; a low P/E does not mean cheap	Shortage largely priced; watch for a 2027 price turn

Two caveats:

• Valuation figures come from different sources and dates; use them to compare relative positions, not as precise values.
• Through a Minervini lens, lagging is itself a signal. A name like Synopsys has weak relative strength; a great structural position does not mean the stock will move. Put it on the watchlist and let its price action prove it.

A test arrives this week: Coherent reports after the close on November 4 and Lumentum after the close on November 5. If the laser shortage narrows clearly, optical names should move down from tier 2.

4. Other sectors

The non-AI part of a portfolio is there to lower correlation with the AI holdings, not to find another high-beta story.

Sector	Direction	Why
Healthcare	Lean positive	Analysts see healthcare benefiting most from Democratic control of both chambers; already among the S&P's strongest sectors before the vote, so partly priced
Power / grid	Neutral to positive	AI power demand and grid investment are long-term trends; tighter permitting may delay some demand, so regulated utilities and grid equipment are steadier than independent power producers
Defense (especially drones)	Wait	Budget risk under a new Congress; AeroVironment and Kratos are already among the biggest decliners
Financials	Wait	Capital requirements and consumer-protection rules may tighten
Cash / short-term bonds	Keep	Cash earns interest in a hiking cycle and is the reserve for when confirmation comes

The policy backdrop: a divided government rarely passes major legislation quickly. Today's reversal is the market pricing "the status quo won't change fast," so sector rotation should not be pushed too far.

5. How to position: discipline matters more than stock picking

By Minervini's approach, this is a time to build a watchlist, not to load up. Five steps:

1. Don't buy into a downtrend. Wait for a follow-through day on the Nasdaq or the Philadelphia Semiconductor Index before starting to buy.
2. Find the stocks that held up best in the pullback. With the index down 20%+, the stocks only about 10% off their highs with intact RS lines are often the next leaders.
3. Build in stages. Start with a quarter position; add to half and full only when it is working. Stop losses at 7–8% per trade.
4. Cross-check with insider activity. After a big drop, whether chip executives are buying in the open market with their own money is a strong confirming signal.
5. Keep enough cash. In a hiking cycle, cash is not wasted; it is the reserve for when confirmation comes.

6. What to watch next

Two catalysts decide when this pullback ends: Nvidia's next earnings report and the December FOMC.

• Nvidia earnings: tests whether the spending can pay back. Watch the data-center revenue outlook and the wording on supply constraints.
• December FOMC: sets the valuation ceiling. If the Fed signals a pause after hiking, pressure on growth stocks eases considerably.
• Cloud providers' cash flow: steady spending is fine; what matters is whether cloud revenue growth keeps up.
• State data-center policy: especially permitting and power-price policy after the new governors take office in Texas and Georgia.

This is a personal research note and thinking framework, not investment advice. Data as of November 4, 2026. Investing involves risk; please make your own judgment.

Sources

• AI data center stocks drop as Democrats sweep U.S. midterms – Investing.com
• How Markets Reacted to Democrats’ Win – Bloomberg
• Midterm elections could jolt stocks – CNBC
• Chip stocks lose $1 trillion as Intel sinks – Benzinga
• Intel’s worst drop since 1974 – Value Add VC
• VIX whipsaw on sticky inflation and AI capex – Yahoo Finance
• Alphabet sheds $500 billion in a single day despite cloud growth of 103% – Motley Fool
• Alphabet Q3 2026: worst day on record explained – Pomegra
• Big Tech earnings: Azure +48%, Alphabet -12.6% – Shattered
• Azure hits 48%, hyperscaler capex outlook – Tech Times
• Alphabet, Microsoft, Meta earnings will judge AI capex – AI Weekly
• Fed hikes, hints at more – InvestingLive
• Why health care stocks are rallying – Axios
• AI policy fears grip traders before midterms – Advisor Perspectives
• CoWoS supply-demand gap seen narrowing – TrendForce
• TSMC's CoWoS bottleneck explained – INDmoney
• TSMC CoWoS packaging bottleneck – Live in the Future
• AI demand keeps 3nm and CoWoS tight through 2027 – Pulse
• TSMC to quadruple advanced packaging capacity – TokenRing
• ASML projects $71B revenue by 2030 – Tom's Hardware
• ASML Q4 bookings double – Rallies
• Nasdaq-100 Inside the Index: ASML – Nasdaq
• EDA Market Primer – SemiAnalysis
• Synopsys: how the business works – Belcourt
• The state of ABF substrates in 2026 – Tom's Hardware
• The ABF substrate bottleneck – Wing VC
• Ajinomoto to raise ABF film prices – OnMSFT
• Substrate shortages and Unimicron expansion – T.Media
• Micron HBM storyline – Silicon Analysts
• KLA: how the business works – Belcourt
• Can process control intensity drive KLA's growth – Nasdaq
• InP shortage emerges as AI optical bottleneck – TrendForce
• The indium phosphide squeeze – Abhay Terminal
• Lumentum touts AI optics boom – MarketBeat
• Lumentum InP AI optics shortage – XenoSpectrum
• The gas turbine shortage just became AI's biggest constraint – OilPrice
• US power firms scramble for transformers and turbines – The Next Web
• NVIDIA vs Broadcom – Metavert
• If AI spending slows in 2027 – 24/7 Wall St.
• Lam Research FY27 Q1 analyst views – StockHub
• Vertiv Q3 2026 earnings call – The Singju Post
• Semiconductor stocks are cheaper than the S&P 500 – Benzinga
• My top 3 AI stock picks for November – Motley Fool
• Semiconductor stocks theme page – StockTitan
• Why I keep buying Taiwan Semiconductor – Yahoo Finance
• What just happened to Synopsys stock – Trefis
• Synopsys valuation and news – Parqet
• Coherent and Lumentum initiated at Buy – TIKR
• Coherent stock fell 11% – TIKR
• Lumentum quote and news – StockAnalysis
• Lumentum forward P/E – StockAnalysis