Semiconductors

A bilingual research brief for The Mid-2026 Semiconductor Correction: AI Infrastructure, Model Mis-Specification, and Memory as a Bottleneck Asset — SSRN Working Paper No. 7013899, Huang, Po-Sung (Sinclair), 2026. What the paper studies. The mid-2026 semiconductor correction generated two competing narratives: a “bubble-burst” reading grounded in legacy DRAM cycle models, and an “infrastructure repricing” reading grounded in bottleneck economics and AI capex fundamentals. The paper uses the divergence in analyst and retail commentary around Micron Technology as a case study in regime mis-specification — the application of a prior-cycle statistical model to a structurally transformed market environment. ...

AI will be everywhere. The harder question is who still earns a margin when intelligence gets cheap. By Sinclair I have watched AI move from conversation and search into data processing, automation, vehicles, robots, industrial tools, and factories. That makes me sceptical of debates that begin with whether AI demand will exist. Much of the eventual demand will not look like a deliberate purchase of “AI.” Intelligence will be built into a product or workflow, and the customer will pay for the outcome. ...

本文為繁體中文版。English version: The AI Capex Money Map v0.2 — America Spends. Who Actually Keeps the Margin? 本站為文字版；含完整圖表的版本請見 Medium 原文。 大多數人現在都在問：「AI 是不是泡沫？」 這個問題重要，但不夠可操作。對投資人與供應鏈觀察者來說，更好的問題只有一句： 美國大型科技公司花出去的 AI Capex，最後流到誰的收入、誰的毛利、誰的護城河？ 這篇不是拆到每一根電纜的投行 BOM 模型，而是一張用公開資料建立的 v0.2 錢流地圖。它的目的不是假裝精準，而是把三件常被混在一起的事情分開來看：錢在哪裡變成營收、營收在哪裡變成毛利、瓶頸何時鬆解。 因為這三題的答案，往往不是同一批公司。 一、AI Capex 不是一個數字，是一組水管 Reuters 引述 Bridgewater 的估計：Alphabet、Amazon、Meta、Microsoft 四大 2026 年的 AI 基礎設施投資約 US$650B，高於 2025 年約 US$410B。先講清楚一件事——這是四大的地板數，不含 Oracle、Stargate、CoreWeave 這類 neocloud、xAI，也不含任何非美系業者。所以 $650B 是四家公司的下限，不是市場總額。 錢不會平均落下，而會進入性質不同的系統層支出。設備與材料則屬二階供應商 capex，不計入 hyperscaler 直接支出，以避免重複計算。 關鍵是——設備與材料（ASML、AMAT、TEL、Advantest…）不在這 $650B 裡。 那是台積電、SK Hynix 的資本支出，是二階供應商支出、是另一個分母。把它加進 hyperscaler 支出裡相加，就是 v0.1 犯的、也是很多市場圖表仍在犯的 double counting。 二、把箱子打開：成本重心已經從邏輯移到 HBM + 封裝 把「Compute Systems」這條水管打開，你會看到這一輪最反直覺的結構變化。 ...

繁體中文版：AI Capex 錢流地圖 v0.2｜美國出錢，誰真的收得到毛利？ The AI Capex Money Map v0.2 — America Spends. Who Actually Keeps the Margin? From $650B to HBM, CoWoS and power — mapping who gets paid, who keeps margin, and when the bottlenecks move. (The capex waterfall, cost-stack, bottleneck-clock, and regional margin-capture charts are figures in the original Medium version; this text edition keeps the surrounding reasoning — see Medium for the full visual breakdowns.) Most people are asking whether AI is a bubble. It is an important question, but not a very operational one. ...

From “What’s the next AI stock?” to “Where’s the next bottleneck, who gets paid, and what would prove us wrong?” Sinclair Huang This essay helps answer four practical questions: Which AI themes are real constraints rather than just attractive stories? Who can capture gross margin when a constraint binds? How long might the bottleneck last before capacity, substitution, or efficiency relieves it? What evidence would prove the thesis wrong? The main argument: large AI demand explains why the sector is hot; bottlenecks explain who gets paid; spillovers explain who absorbs hidden costs; monetised usage determines whether the buildout is sustainable. ...

COMPUTEX 2026 and Taiwan’s shift from supply chain to co-design and deployment marketplace AI Infrastructure Notes｜Article 4 Sinclair Huang Field Note v4 — updated with GTC Taipei and statistical-science reference materials, using public and user-provided sources available as of June 6, 2026. I am writing this at a strange distance from Taiwan. While friends in Taipei are celebrating a market that seems to have escaped ordinary scale, I am abroad, watching the same numbers with admiration, unease, and a growing sense that another AI market commentary would not be enough. ...

Product exposure, process-control exposure, and the physical bottlenecks behind the AI capex wave AI Infrastructure Notes | Part 3 Sinclair Huang A reader recently left a comment on my ABF substrate piece that stayed with me. His point was simple: CoWoS and HBM get most of the attention, but substrate materials often sit below the level where many equity models even begin. I think that observation captures a broader problem in AI infrastructure analysis. ...

Why CPO is not an optics story — it is a process-integration story. AI Infrastructure Notes | Part 2 Sinclair Huang Everyone says AI needs more bandwidth. That part is true. As AI clusters scale from thousands of accelerators to tens of thousands, and then toward million-GPU-scale systems, the network stops being a background layer. It becomes part of the compute fabric itself. Copper reaches its distance, power, and signal-integrity limits. Optical interconnect moves closer to the switch ASIC. Co-packaged optics becomes the logical next step. ...

Everyone Is Betting on AI — But Almost No One Asks: Where Are You Standing?The AI boom is real. The harder question is where value, risk, and leverage accumulate inside the stack. Position matters more than opinion. Over the past few weeks, I have been thinking about AI through a different question: Why can the same AI cycle feel so powerful in markets, yet so uncertain in daily life? My own view is more optimistic than pessimistic. ...

From compute bottlenecks to industrial consequences — where value may actually concentrate through 2030 Series: AI Compute Supply Chain | Part 5 of 5 Author: Po-Sung (Sinclair) Huang For the past four articles in this series, I have written about CoWoS, HBM, ABF substrates, SEC filings, and the fault lines that could eventually crack today’s moats. On the surface, that may look like a semiconductor series. It is not. What these articles really reveal is something larger: AI is no longer just a software story, and no longer just a model race. It is becoming an industrial system — one that depends on power, cooling, capital expenditure, advanced packaging, memory bandwidth, substrate materials, qualification cycles, and the physical discipline of manufacturing scale. ...