AI Supply Chain Research

A ten-sector, source-verified research series on the AI compute build-out — from accelerators to gigawatts — plus an independent credibility audit of the source it distills.

📖 Read it live: yichengyang-ethan.github.io/ai-supply-chain · ⚖️ Credibility audit: How Reliable Is SemiAnalysis?

📕 PDF (Public Edition): AI_Supply_Chain_Research_Complete_Public_Edition.pdf — the full 199-page book (English): an 11-module Foundations Primer (no background assumed), the panorama, ten sector deep-dives, and the credibility audit. Per-chapter linked sources.

🎓 New to AI hardware? Start with the primer: live · data in data/primer/ — 11 textbook-style modules, 76/76 key numbers source-verified, own-drawn diagrams (labeled), and attributed institutional images.

By Yicheng Yang · independent, not affiliated with SemiAnalysis · nothing here is investment advice.

---

What this is

The AI build-out is one machine viewed through ten cross-sections: accelerators, foundry, memory/HBM, advanced packaging, semi equipment, networking/optics, datacenter power, hyperscaler economics, models/software, and China/export controls. This repo contains a structured, verified distillation of that machine from 311 public SemiAnalysis articles (2020–2026), organized for investment research:

• ~130 theses across ten sectors — each a claim + rationale + attributed, linked quote
• ~170 quantified data points — every number grep-verified against the underlying text
• Per-sector positioning: winners, bottlenecks, risks, catalysts
• Six cross-sector through-lines — the demand shock, the bottleneck relay, chip→rack→campus co-design, the value-capture war, the software moat, China's bifurcation

Why trust it — the verification layer

This isn't a summary; it's an audited distillation:

Check	Result
Quantitative claims verified against source text	1,070 / 1,091 (98.1%) — the 20 flagged were corrected or removed
Quotes verified verbatim (or ellipsis-stitched from verbatim fragments)	131 / 131 after 3 corrections
Source images reproduced	0 — every chart reference is a link, per the source's attribution terms
Every claim sourced and linked	✅ ~190 article links

The credibility audit (the part you should read first)

Before trusting one source this much, I audited it: an independent event study of 664 stance-bearing company mentions across 119 tickers extracted from the corpus, measuring excess returns vs. SMH after SemiAnalysis praised or panned a name (equal-weight by name, ticker-clustered, pre-IPO look-ahead removed, delistings settled).

Findings (full scorecards in audit/):

1. The only significant leading signal lasts ~1 month — bull-minus-bear spread +3.9% (Welch t = 2.33).
2. Long horizons reverse: the typical bullish name lags SMH by −5.2% at 1Y and −27.2% at 3Y (t = −4.49).
3. Aggregate returns are a beta illusion — excluding the five most-covered names (NVDA/TSM/AMD/GOOGL/AVGO), bullish 1Y excess is −6.1%.
4. Bearish calls beat bullish ones: panned names lag SMH −10.7% at 1Y (t = −2.90). INTC (panned 30×) and Samsung (15×) both played out.
5. Right industry, wrong stock: the memory-shortage / CoWoS / datacenter-power narratives were all right — the money still went to NVDA/TSM.

Takeaway: treat the source as a microscope on mechanisms and bottlenecks, not a buy list. That is exactly what this series distills. (Fairness note: SemiAnalysis explicitly states it does not provide investment advice or ratings — scoring it as a stock-picker is my frame, not theirs. Methodology limits are stated plainly in audit/README.md.)

Repo layout

data/
  primer/           primer_XX.json — 11 Foundations modules (EN + ZH), every number sourced
  sectors/          content_XX.json — ten verified sector briefs (EN + ZH), the canonical data
  synthesis/        cross-sector through-lines, chain-wide investment map, audit narrative
  source_links.json title → SemiAnalysis URL map (~190 articles)
audit/
  bullish_scorecard.csv / bearish_scorecard.csv — per-name post-coverage excess returns
  headline_stats.json — aggregate results with t-stats
  README.md — methodology & limits
tools/
  generate_site.py — deterministic JSON → static-site generator (the live site is built from this)

Reproduce the site

python3 tools/generate_site.py     # writes 12 static pages to ./site_output

No dependencies beyond the Python standard library. The generator enforces the compliance rules (no third-party images; quotes render only if resolvable to a source link).

Method

1. Corpus: 311 publicly accessible SemiAnalysis articles (free posts + free previews; no paywall circumvention).
2. Distillation: parallel LLM agents read each sector's full bucket and emit structured briefs (theses, data, deep-dives) in both English and Chinese.
3. Adversarial verification: independent fact-checking agents grep every number and quote against the corpus; fabrications/distortions are corrected by prescription. Mechanical checks validate all quotes and link mappings.
4. Credibility audit: stance extraction (27 agents) → ticker resolution (incl. KRX/TWSE/TSE/HKEX) → event study vs. SMH with the debiasing steps above, after three adversarial audit rounds fixed eight methodological defects.

License & attribution

• Code (tools/): MIT. Data files in data/ and audit/: CC BY 4.0.
• All references to SemiAnalysis research are quoted/excerpted with attribution and links, per their published attribution terms. This project is independent and unaffiliated; for the full research, subscribe to SemiAnalysis.