AI-SEMI 美股 AI 半导体产业链深度调研
K线走势 + 关键信号位Candles + Key Signal Levels
暂无可视化价格信号。No visual price signals provided.
一屏速读Quick Read
- 先看层级：最先卡住的是能否把 AI 芯片、HBM、封装和网络规模化交付，而不是谁 PPT 里 AI 故事讲得最响。Layer first: the scarce part is whether AI chips, HBM, packaging, and networking can scale into real shipments, not who tells the loudest AI story.
- AVGO 排第一，是因为它同时卡住定制 XPU 和 AI Ethernet 网络；TSM 排第二，是因为几乎所有高端 AI 芯片最终都绕不开先进制程和 CoWoS。AVGO ranks first because it sits at both custom XPU and AI Ethernet networking. TSM ranks second because high-end AI chips still depend on advanced process and CoWoS.
- MU 是 HBM 稀缺层里最直接的美股纯标的；MRVL 更偏 AI 连接弹性；KLAC 是不直接卖 AI 芯片、但卡住先进制程和先进封装良率的工具层。MU is the direct U.S.-listed HBM scarcity play; MRVL is higher-beta AI connectivity; KLAC is the tool layer that bottlenecks advanced-node and packaging yield.
分批进场点位与仓位计划Staged Entry Levels & Position Plan
| 触发条件Trigger | 进场点位Entry | 单次仓位Position | 止损/失效Stop | 观察目标Target | 执行说明Execution Note |
|---|---|---|---|---|---|
| 优先研究 #1：AVGOPriority #1: AVGO | 定制 XPU / AI Ethernet定制 XPU / AI Ethernet | 研究权重 25%研究权重 25% | AI 收入指引放缓或大客户订单不透明AI 收入指引放缓或大客户订单不透明 | 财报指引 / Meta、OpenAI、云客户公告财报指引 / Meta、OpenAI、云客户公告 | 最接近客户自研芯片和网络交换双卡点。Closest to both custom silicon and AI networking bottlenecks. |
| 优先研究 #2：TSMPriority #2: TSM | 先进制程 / CoWoS先进制程 / CoWoS | 研究权重 22%研究权重 22% | CoWoS 扩产缓慢、地缘/水电/专利风险升温CoWoS 扩产缓慢、地缘/水电/专利风险升温 | 20-F、月营收、CoWoS 扩产、Arizona 项目20-F、月营收、CoWoS 扩产、Arizona 项目 | 整个 AI 芯片链的“总闸门”。The master gatekeeper for high-end AI chips. |
| 优先研究 #3：MUPriority #3: MU | HBM / 高端 DRAMHBM / 高端 DRAM | 研究权重 20%研究权重 20% | HBM 认证/良率不及预期或内存周期反转HBM 认证/良率不及预期或内存周期反转 | HBM4、NVIDIA 认证、CHIPS/环保许可HBM4、NVIDIA 认证、CHIPS/环保许可 | 美股里最直接的 HBM 稀缺标的。The most direct U.S.-listed HBM scarcity play. |
| 优先研究 #4：MRVLPriority #4: MRVL | 光 DSP / 1.6T / 102.4T Switch光 DSP / 1.6T / 102.4T Switch | 研究权重 18%研究权重 18% | 客户集中、定制 ASIC 能见度下降客户集中、定制 ASIC 能见度下降 | OFC、102.4T 交换、数据中心收入OFC、102.4T 交换、数据中心收入 | 弹性高，但商业可见度和估值风险也更高。High upside beta, but with higher visibility and valuation risk. |
| 优先研究 #5：KLACPriority #5: KLAC | 过程控制 / 先进封装良率过程控制 / 先进封装良率 | 研究权重 15%研究权重 15% | WFE 周期下行或估值过高WFE 周期下行或估值过高 | 过程控制强度、先进封装检测、毛利率过程控制强度、先进封装检测、毛利率 | 卖铲子里最接近良率卡点的标的。The picks-and-shovels name closest to yield bottlenecks. |
超买超卖 + RSI + 布林带Overbought/Oversold + RSI + Bollinger Bands
当前输出为产业链排序和标的研究优先级。Current output is layer ranking and ticker research priority.
不适用N/A · 本报告是产业链深度调研，不针对单一图表做超买超卖判断。This is supply-chain research, not a single-chart overbought/oversold analysis.
若后续指定单一标的，可再加入 TradingView 图表分析。For a single ticker, TradingView chart analysis can be added later.
| 指标Metric | 数值/状态Value/Status | 分析Analysis |
|---|---|---|
| 超买/超卖状态Overbought/Oversold | 不适用N/A | 当前输出为产业链排序和标的研究优先级。Current output is layer ranking and ticker research priority. |
| RSIRSI | N/AN/A | 不适用N/A |
| 布林带位置Bollinger Position | 不适用N/A | 若后续指定单一标的，可再加入 TradingView 图表分析。For a single ticker, TradingView chart analysis can be added later. |
| 布林带数值Bollinger Values | U N/A / M N/A / L N/AU N/A / M N/A / L N/A | 上轨/中轨/下轨Upper / Middle / Lower |
7个维度评分拆解Seven-Dimension Score Breakdown
AI 训练/推理、定制 ASIC、HBM 和数据中心网络仍处强周期。AI training/inference, custom ASICs, HBM, and data-center networking remain in a strong cycle.
最强壁垒在 CoWoS、HBM、SerDes/DSP、EDA/IP、过程控制和客户认证。Strongest moats sit in CoWoS, HBM, SerDes/DSP, EDA/IP, process control, and customer qualification.
AVGO/MU/KLAC/AMAT 等最新财报显示现金流和利润率强，但估值分化大。Latest results for AVGO/MU/KLAC/AMAT show strong profitability/cash flow, but valuation dispersion is large.
短期确定性来自真实订单、产能预订、CHIPS 项目和客户认证。Near-term certainty comes from real orders, capacity reservations, CHIPS projects, and customer qualifications.
半导体 AI 主线拥挤，很多标的已经预支 2027 增长。AI semiconductor trade is crowded, with many names pre-pricing 2027 growth.
下一轮催化看财报指引、HBM 认证、CoWoS 扩产、网络交换新品和客户公告。Next catalysts are earnings guidance, HBM qualification, CoWoS expansion, networking launches, and customer announcements.
风险来自估值杀、客户集中、出口管制、CapEx 兑现不及预期。Risks are valuation compression, customer concentration, export controls, and AI capex disappointment.
Serenity产业链卡点Serenity Scarce-Layer Lens
市场故事Market Story: AI 算力需求继续扩张，但市场已经从“谁有 GPU”转到“谁能把 GPU/ASIC/HBM/网络/封装真正规模化交付”。AI compute demand keeps expanding, but the market has shifted from who has GPUs to who can actually scale GPU/ASIC/HBM/networking/packaging shipments.
系统变化System Change: 大模型训练和推理从单芯片性能竞争，进入整机柜、集群网络、内存带宽、功耗、封装良率和交付周期竞争。AI competition has moved from single-chip performance into rack-scale systems, cluster networks, memory bandwidth, power, packaging yield, and delivery cycles.
公司位置Target Position: 本次排序故意没有把 NVDA 放入前五，因为 NVDA 是主线锚，不是当前最适合找“市场可能没看清”的卡点；它更适合作为验证整条链需求的温度计。NVDA is intentionally not in the top five: it is the anchor, not the best place to find underappreciated bottlenecks. It is better used as the thermometer for the whole chain.
风险警戒与退出逻辑Risk Alerts & Exit Logic
趋势Trend: AI 半导体主线仍强，但估值和预期拥挤。AI semiconductor theme remains strong, but valuations and expectations are crowded.
支撑Support:
压力Resistance:
突破确认Confirmation:
失效止损Invalidation:
稀缺层级排序Scarce Layer Ranking
| 层级Layer | 为什么重要Why It Matters |
|---|---|
| 1. 先进封装 / CoWoS / HBM 集成1. Advanced packaging / CoWoS / HBM integration | 没有先进封装和 HBM 集成，高端 GPU/ASIC 无法真正出货。TSMC CoWoS 官方定位就是 AI/HPC 的异构集成平台。Without advanced packaging and HBM integration, high-end GPU/ASIC shipments cannot scale. TSMC positions CoWoS as an AI/HPC heterogeneous integration platform. |
| 2. HBM / 高端 DRAM2. HBM / high-end DRAM | HBM 决定每颗加速器的数据吞吐和供给上限；Micron 已开始 HBM4 出货/推进高层数产品。HBM determines data throughput and supply ceiling per accelerator; Micron has started HBM4 shipment/ramp and higher-stack products. |
| 3. 定制 XPU + AI 以太网交换3. Custom XPU + AI Ethernet switching | 云厂商自研芯片需要定制硅和网络平台，Broadcom/Marvell 卡在这里。Hyperscaler custom silicon needs custom ASIC and networking platforms, where Broadcom and Marvell sit. |
| 4. 光互连 DSP / 1.6T / CPO4. Optical DSP / 1.6T / CPO | 集群规模扩大后，铜互连距离和功耗成为限制，光 DSP 变成 AI fabrics 的关键。As clusters scale, copper reach and power become limiting, making optical DSP critical for AI fabrics. |
| 5. 过程控制 / 量测 / 良率5. Process control / metrology / yield | 先进封装和先进制程复杂度提升后，每一点良率都影响有效产能，KLA/AMAT 受益。As advanced packaging and nodes grow complex, each point of yield affects effective capacity; KLA/AMAT benefit. |
业务结构Business Structure
| 业务Segment | 收入占比Revenue Share | 增长Growth | 说明Note |
|---|---|---|---|
| 先进封装/CoWoSAdvanced packaging / CoWoS | 层级排名 #1Layer rank #1 | 极紧Extremely tight | TSM 是总闸门；AMKR/ASE 等是扩容观察对象。TSM is the master gatekeeper; AMKR/ASE are capacity-expansion watch names. |
| HBM / DRAMHBM / DRAM | 层级排名 #2Layer rank #2 | 供需紧Tight supply/demand | MU 是美股最直接标的，SK hynix/Samsung 是全球对照。MU is the direct U.S.-listed play; SK hynix/Samsung are global comps. |
| 定制 XPU / AI 网络Custom XPU / AI networking | 层级排名 #3Layer rank #3 | 高High | AVGO 与 MRVL 是主要美股研究线。AVGO and MRVL are the main U.S.-listed research lines. |
| 过程控制/设备Process control / equipment | 层级排名 #4Layer rank #4 | 强但估值高Strong but expensive | KLAC 更偏良率，AMAT 更偏广谱设备平台。KLAC is yield-control heavy; AMAT is broader WFE platform. |
毛利与盈利质量Margin & Profit Quality
| 指标Metric | 数值Value | 趋势Trend | 说明Note |
|---|---|---|---|
| 财务质量最佳Best financial quality | AVGO / KLAC / AMATAVGO / KLAC / AMAT | 高利润率High margins | AVGO adjusted EBITDA margin 68%；KLA 指引非 GAAP 毛利率约 61.75%；AMAT 半导体系统毛利率约 54.8%。AVGO adjusted EBITDA margin 68%; KLA guides non-GAAP GM around 61.75%; AMAT semiconductor systems GM around 54.8%. |
| 增长爆发最强Strongest growth burst | MU / AVGO / MRVLMU / AVGO / MRVL | AI 拉动AI driven | MU Q2 FY2026 收入同比大幅增长；AVGO AI 收入同比 +106%；MRVL 数据中心和光互连验证强。MU Q2 FY2026 revenue surged; AVGO AI revenue +106%; MRVL data center/optical connectivity evidence is strong. |
| 最大估值风险Highest valuation risk | MRVL / KLAC / AVGOMRVL / KLAC / AVGO | 预期拥挤Crowded expectations | 如果云 CapEx 或 AI 指引低于预期，估值杀会先打高弹性标的。If cloud capex or AI guidance disappoints, high-beta names face valuation compression first. |
财务指标Fundamentals
| 指标Metric | 数值Value | 说明Note |
|---|---|---|
| AVGOAVGO | AI revenue $8.4B, +106% YoY in Q1 FY2026AI revenue $8.4B, +106% YoY in Q1 FY2026 | 定制加速器 + AI networking 是核心。Custom accelerators + AI networking are core. |
| MUMU | Q2 FY2026 revenue $23.86B; operating cash flow $11.90BQ2 FY2026 revenue $23.86B; operating cash flow $11.90B | HBM/高端内存周期验证强。HBM/high-end memory cycle validation is strong. |
| KLAKLA | Q4 FY2026 revenue guide $3.575B +/- $200MQ4 FY2026 revenue guide $3.575B +/- $200M | 非 GAAP 毛利率指引约 61.75% +/- 1%。Non-GAAP gross margin guide around 61.75% +/- 1%. |
| AMATAMAT | Q2 FY2026 revenue $7.91B, +11% YoYQ2 FY2026 revenue $7.91B, +11% YoY | 设备链财务质量强，但未进前五是因为卡点不如 KLAC 精准。Strong equipment financials, but less bottleneck-specific than KLAC. |
多智能体博弈投票Six-Agent Debate
结论偏多，但只看真正卡住供给的层级；买点要等估值和财报验证。Consensus is constructive, but only for true supply bottlenecks; entries require valuation and earnings validation.
AI 半导体仍是最强主线，但热门 GPU 和高弹性票已经很拥挤。AI semis remain the strongest theme, but GPUs and high-beta names are crowded.
CoWoS、HBM、SerDes/DSP、过程控制比泛 AI 故事更值得研究。CoWoS, HBM, SerDes/DSP, and process control are better research targets than generic AI stories.
估值和预期都高，任何财报指引瑕疵都会放大回撤。Valuations and expectations are high; any guidance flaw can amplify drawdowns.
板块涨幅大，研究优先级不等于追高买点。The sector has run; research priority is not a chase entry.
高毛利、高现金流、真实订单比概念更重要。High margins, cash flow, and real orders matter more than concept.
若 AI 收益率被质疑，整条链可能一起杀估值。If AI ROI is questioned, the whole chain can derate together.
同行对比Peer Alternatives
| 代码Ticker | 位置Position | 相对观点Relative View |
|---|---|---|
| AVGOAVGO | 定制 XPU / AI Ethernet 网络Custom XPU / AI Ethernet networking | #1。卡住大客户自研芯片和集群网络，AI 收入实锤，财务质量强。#1. Bottlenecks hyperscaler custom silicon and cluster networking, with proven AI revenue and strong financial quality. |
| TSMTSM | 先进制程 / CoWoS / AI 芯片总闸门Advanced process / CoWoS / AI chip gatekeeper | #2。几乎所有高端 AI 芯片绕不开 TSM，但地缘、水电、客户集中风险要折价。#2. Nearly all high-end AI chips depend on TSM, but geopolitics, water/power, and customer concentration deserve a discount. |
| MUMU | HBM / 高端 DRAMHBM / high-end DRAM | #3。最直接的美股 HBM 稀缺标的，需盯 NVIDIA 认证、良率和内存周期。#3. The most direct U.S.-listed HBM scarcity play; watch NVIDIA qualification, yields, and memory cycle risk. |
| MRVLMRVL | 光 DSP / 1.6T / AI 交换 / 定制硅Optical DSP / 1.6T / AI switching / custom silicon | #4。弹性强，卡住连接层；主要风险是客户集中和商业可见度。#4. High beta and close to the connectivity bottleneck; main risks are customer concentration and visibility. |
| KLACKLAC | 先进封装/先进制程过程控制与良率Advanced packaging/node process control and yield | #5。不直接卖 AI 芯片，但卡住良率；财务质量强，估值也贵。#5. Does not sell AI chips directly, but bottlenecks yield; high quality, expensive valuation. |
风险清单Risk List
| 风险Risk | 等级Severity | 观察条件Watch |
|---|---|---|
| AI CapEx 兑现不及预期AI capex disappointment | 高High | 云厂商财报、GPU/ASIC 订单、数据中心电力项目延迟。Cloud earnings, GPU/ASIC orders, data-center power project delays. |
| CoWoS/HBM 扩产太快导致价格回落CoWoS/HBM oversupply after rapid expansion | 中高Medium-High | HBM 长协价格、Micron/SK hynix/Samsung capex、TSMC packaging lead time。HBM LTAs, Micron/SK hynix/Samsung capex, TSMC packaging lead times. |
| 出口管制和地缘风险Export controls and geopolitics | 高High | 美国对中国 AI 芯片/设备限制、台湾风险、ITC/专利纠纷。U.S. China AI chip/equipment restrictions, Taiwan risk, ITC/patent disputes. |
| 客户集中与定制芯片项目取消Customer concentration and custom-chip cancellation | 高High | AVGO/MRVL 的 hyperscaler 项目节奏与收入披露。AVGO/MRVL hyperscaler project timing and revenue disclosure. |
数据完整性与来源Data Integrity & Sources
| 结论/事实Claim | 来源类型Source Type | 证据强度Strength |
|---|---|---|
| Broadcom Q1 FY2026 AI revenue 84亿美元，同比 +106%，由 custom AI accelerators 和 AI networking 驱动。Broadcom Q1 FY2026 AI revenue was $8.4B, +106% YoY, driven by custom AI accelerators and AI networking. | Broadcom IR 财报公告Broadcom IR earnings release | 强Strong |
| TSMC CoWoS 官方说明其用于 AI/HPC 异构集成，可容纳逻辑 chiplets 与 HBM。TSMC's CoWoS documentation states it supports AI/HPC heterogeneous integration with logic chiplets and HBM. | TSMC 3DFabric/CoWoS 官方资料TSMC 3DFabric/CoWoS official material | 强Strong |
| Micron Q2 FY2026 收入 238.6亿美元，经营现金流 119亿美元，HBM4 已开始出货/推进。Micron Q2 FY2026 revenue was $23.86B, operating cash flow $11.90B, and HBM4 shipment/ramp is underway. | Micron IR 财报与演示Micron IR earnings and presentation | 强Strong |
| Marvell 官方说明 AI 负载推动 800G 到 1.6T 光互连，DSP 是低延迟、低功耗光链路关键。Marvell states AI workloads drive 800G to 1.6T optical interconnect and DSPs are key for low-latency, low-power optical links. | Marvell optical DSP 官方资料Marvell optical DSP official material | 强Strong |
| KLA 官方资料显示其先进封装组合覆盖晶圆、面板和组件过程控制，帮助提升封装良率。KLA's official material shows advanced packaging process control across wafers, panels, and components to improve yield. | KLA Advanced Packaging 官方资料KLA Advanced Packaging official material | 强Strong |
| TSMC Arizona 获 CHIPS 直接资金最高 66亿美元；Micron 获美国本土内存制造相关 CHIPS 支持，且 NYSDEC 于 2026-03-31 签发 Micron Air Title V permit。TSMC Arizona received up to $6.6B CHIPS direct funding; Micron received U.S. memory manufacturing support and NYSDEC issued Micron's Air Title V permit on 2026-03-31. | NIST / NYSDEC / 公司公告NIST / NYSDEC / company releases | 强Strong |
下一步观察Next Checks
- 逐一读 AVGO、MRVL、MU、TSM、KLAC 的最新 10-Q/10-K/20-F，重点看客户集中、订单能见度、capex、库存和毛利率。Read latest AVGO/MRVL/MU/TSM/KLAC 10-Q/10-K/20-F; focus on customer concentration, order visibility, capex, inventory, and margins.
- 跟踪 CoWoS 交期、HBM 合同/认证、NVIDIA/AMD/云厂商 rack 规格变化。Track CoWoS lead times, HBM contracts/qualifications, and NVIDIA/AMD/hyperscaler rack spec changes.
- 查 SEC comment letters/ITC 专利纠纷/出口管制更新；目前美股没有互动易这种同形态渠道。Check SEC comment letters, ITC patent disputes, and export-control updates; U.S. names do not have an Eastmoney Hudongyi equivalent.
- 对每个标的再做单票图表和估值安全边际，研究优先级不等于立即买点。Run single-name chart and valuation work for each candidate; research priority is not immediate entry.
免责声明Disclaimer
仅供公开资料研究参考，不构成投资建议或交易指令。美股半导体波动大，排序代表研究优先级，不代表买卖建议。Research support based on public information only; not investment advice or trading instruction. Ranking means research priority, not a buy/sell recommendation.