Compute supply chains, deployment constraints, and where capital spending becomes durable margin.
AI infrastructure is not one trade. It is a chain of moving constraints: compute, memory bandwidth, advanced packaging, power, cooling, networking, qualification, and deployment. The capex wave is real, but spending does not become durable profit simply because it carries an AI label. Value accrues where a supplier controls a bottleneck customers cannot quickly bypass. The key is how long that constraint survives capacity expansion, substitution, and price compression. This research follows the buildout from U.S. capex announcements to Taiwan’s engineering queues, asking four questions: What is physically scarce? On what clock can it expand? Who carries the execution and financing burden? Who still keeps margin when scarcity eases? The aim is not to predict the next AI stock. It is to distinguish deployable infrastructure from narrative exposure, temporary shortage rents from structural value capture, and market repricing from demand collapse.
中文｜核心觀點
AI 基礎設施不是單一題材，而是一連串會移動的瓶頸。需求與資本支出都是真的，但營收流入不等於利潤留存。關鍵在於：什麼限制部署、多久才能擴產或替代、誰承擔整合與資金壓力，以及稀缺性消退後誰仍能保住利潤。這組研究從台灣工程現場出發，追蹤算力、記憶體、封裝、電力、散熱到實際部署。
Start Here — Reading Path
Read these in sequence: from the constraint map, to due diligence, capital flow, deployment, and economic durability.
-
AI Infrastructure Is Not One Trade — Separate product exposure, process control, and infrastructure exposure before treating AI capex as a single theme.
-
Everyone Wants the Next AI Stock. I’m Looking for the Next Physical Bottleneck. — Turn the thesis into a testable discipline: map demand, constraint severity, margin capture, duration, and the evidence that would invalidate the story.
-
The AI Capex Money Map v0.2 — America Spends. Who Actually Keeps the Margin? — Follow hyperscaler spending through revenue, margin, and bottleneck clocks to see where capital may actually settle.
-
AI Needs a Place to Land — See why compute becomes economic value only after power, cooling, verification, integration, and deployment come together.
-
Everyone Is Counting Tokens. Watch the Bandwidth That Gets Paid — Finish with the durability test: whether paid usage can outrun falling unit economics as each scarcity premium decays.
Papers
-
Featured research — The Mid-2026 Semiconductor Correction: AI Infrastructure, Model Mis-Specification, and Memory as a Bottleneck Asset — Applies the author’s IRI, ECDR, and gross-margin framework to ask whether the mid-2026 correction reflected demand collapse or macrofinancial repricing; it is a stress test, not a price forecast.
-
Deeper research — Infrastructure-Led Leading Indicators in Technology Investment Cycles: Evidence from the Semiconductor Industry — Tests whether equipment-sector revenue leads semiconductor capital expenditure and introduces ECDR as a framework for monitoring cycle stress.
All Posts in This Cluster
- Everyone Is Counting Tokens. Watch the Bandwidth That Gets Paid
- The AI Capex Money Map v0.2 — America Spends. Who Actually Keeps the Margin?
- AI 需求沒有消失，但價格開始接受壓力測試
- Everyone Wants the Next AI Stock. I’m Looking for the Next Physical Bottleneck.
- AI Needs a Place to Land
- AI Infrastructure Is Not One Trade
- Beyond the GPU: What the AI Infrastructure Buildout Means for the Real Economy
- The Bottleneck Nobody Is Pricing In: Where AI Compute Really Breaks
Working on These Questions?
Facing a specific AI-infrastructure, supply-chain, or capital-allocation question? I undertake defined-scope project engagements for boards, executive teams, and investors.