P3-C3 · Real Process for Finding Industry Bottlenecks¶
Core One-Liner
Industry bottlenecks precede stock price movements by 6 months — if you see it, you make big gains; if you don't, you're late to the game.
Real Analysis Process — Hedge Fund / Buffett / Finding Industry Bottlenecks / Multi-PM / Anti-thesis
P3-C3 (Part 3, Chapter 3 of 5). After this chapter, you'll be able to use a 5-step process to find the next AI industry bottleneck, 6 months before the stock price reacts.
1. The Problem: Your Thesis Follows the News — By the Time News Breaks, the Stock Has Already Moved¶
Retail investor thesis pain points:
- Seeing "SK Hynix +130% YoY" — already priced in
- Seeing "Constellation Energy up 200%" — MSFT PPA already signed
- Seeing "Stargate $500B" — ORCL / VRT / CEG already rallied 7 days after announcement
→ News is a lagging signal. You want to capture leading signals = industry bottlenecks that appeared 6 months before the news.
Real case: HBM shortage was visible in public data as early as H1 2023, but SK Hynix's stock surged in Q4 2023–Q1 2024. If you had used the 5-step process to spot the bottleneck 6 months earlier, you would have bought in advance. This is the real gap between institutions (Coatue / Druckenmiller) and retail investors.
2. The Solution: 5-Step Bottleneck-Finding Process¶
| Step | What to Do | Leading Time |
|---|---|---|
| 1. Identify bottleneck candidate | What physical resource constrains AI growth | Continuous monitoring |
| 2. Track supply | Capacity / expansion plans / lead time for that resource | Quarterly (earnings) |
| 3. Track demand signal | Hyperscaler capex guide / customer lock-ins | Quarterly |
| 4. Track price / lead time | Price increases + lead time extension = real shortage | Monthly (news) |
| 5. Track sentiment / stock price | Sell-side upgrades / 13F accumulation | Monthly–Quarterly (lag) |
The 5 steps have a time gradient — the earlier you spot the signal, the more alpha you capture.
Key insight: Steps 1–2 use public data anyone can access, but 99% of retail investors don't bother. Reading company IR reports + earnings transcripts already puts you ahead of 90%.
3. How It Works: HBM Shortage 2023–2024 Real Timeline¶
3.1 Step 1 — Identify Bottleneck Candidate (Q1 2023)¶
LLM training requires HBM (High Bandwidth Memory, covered in P1-C5). Physical bottleneck:
- Only 3 companies make HBM (SK Hynix / Micron / Samsung)
- 3D stacking is difficult; capacity expansion takes 2 years (fab construction + yield ramp)
- NVDA H100 / H200 shipment ceiling = HBM capacity ceiling
How to spot this as a bottleneck (Q1 2023): - Read NVDA earnings call: Jensen mentioned "supply constrained on HBM" - Read SK Hynix earnings: HBM backlog doubled - → Any company earnings keyword "supply constrained" / "long lead time" is a bottleneck signal
→ What retail can do: Force yourself to read quarterly transcripts from 4 companies (NVDA / SK Hynix / Micron / ASML) and monitor keywords.
3.2 Step 2 — Track Supply (Q2 2023)¶
SK Hynix publicly disclosed: - HBM3e mass production timeline (H1 2024) - Capacity expansion capex (+$5B+ annually) - Customer commitments (NVDA locked orders in December)
Micron publicly disclosed: - HBM3e mass production timeline (H2 2024, later than SK Hynix) - Slow yield ramp
Samsung publicly disclosed: - HBM3 qualification with NVDA in progress, failed - Strike disruptions
→ What retail can do: Read earnings reports from all 3 HBM suppliers for at least 2 quarters, calculate total HBM capacity vs NVDA + AMD GPU demand.
3.3 Step 3 — Track Demand (Q3 2023)¶
Hyperscaler capex guidance (Q3 2023 earnings): - MSFT FY24 capex guide: $44B → $55B (+25%) - META: $35B → $40B - AMZN: $50B → $60B - GOOGL: $35B → $45B
Total ~$200B, almost all AI infrastructure. GPU demand = HBM demand.
Simple math: $200B capex × 50% GPU = $100B GPU revenue, at $30K per GPU = 3.3M GPUs. Each GPU has 96GB HBM (H100) → 300M GB HBM demand (annual). Global HBM capacity at the time was ~150M GB.
→ Demand / Supply = 2x. Shortage confirmed.
→ What retail can do: The calculation above uses free data (company IR / Yahoo Finance) and takes 15 minutes.
3.4 Step 4 — Track Price + Lead Time (Q4 2023)¶
Public signals: - SK Hynix HBM3e price increase (free articles from various industry media) - NVDA H100 lead time extended to 36 weeks (publicly acknowledged in CRWV interview) - Samsung HBM hit by both strike and qualification failure
→ Shortage already visible in public data.
3.5 Step 5 — Sentiment Follow-Up (Q1–Q2 2024)¶
- Sell-side upgrades: multiple analysts raised SK Hynix / Micron price targets
- 13F accumulation: Q1 2024 13F (disclosed 5/15) showed hedge funds adding SK Hynix / Micron
- Stock catch-up: SK Hynix +60% YTD in Q1
→ This is when retail sees the news. But the leading edge (Steps 1–4) appeared 9–12 months earlier.
4. vs P3-C2 — What You Already Know¶
| Dimension | P3-C2 Gives You | P3-C3 Adds |
|---|---|---|
| Buffett 5-Step | Company-level (moat / price) | Doesn't cover industry bottlenecks |
| Finding bottlenecks | ✗ | 5-step process + time gradient |
| Time lead | Follows the news | Leading by 6–12 months |
P3-C2 = company-level bottom-up. P3-C3 = industry-level top-down. Combine both — identify (a) industry bottleneck (b) beneficiary companies © filter with Buffett 5-step → high probability of winning.
5. Try It: Find the Next Industry Bottleneck¶
Task (~1 hour): Use the 5-step process to find the next AI bottleneck for 2026–2027.
Candidates (pick one to deep-dive):
| Candidate Bottleneck | Early Signals | Beneficiary Candidates |
|---|---|---|
| Power (1 GW/data center) | Earnings keywords: "power-constrained" | CEG / VST / GEV |
| HBM4 (mass production 2026) | SK Hynix / Micron 2025 earnings ramp timeline | SK Hynix / Micron |
| CPO (Co-Packaged Optics) | TSM / AVGO / Coherent 2025 earnings | TSM / Coherent / LITE |
| Liquid cooling penetration (B200 1200W) | Vertiv backlog | VRT / Equinix |
| Foundry 2nm (mass production 2025) | TSM / Samsung qualification timing | TSM / Samsung (long-term) |
5-Step Task:
| Step | You Can |
|---|---|
| 1. Physical bottleneck + supplier oligopoly? | y/n |
| 2. Read supplier earnings for 4 quarters, list capacity | ___ |
| 3. Calculate demand (derive from hyperscaler capex) | ___ |
| 4. Find public price / lead time signals | ___ |
| 5. Check 13F / sell-side sentiment | ___ |
Self-check (3 items checked → proceed to P3-C4):
- You can list 5 potential AI industry bottleneck candidates for the next 12 months
- You can monitor 1 of them using free sources (SEC / company IR / public podcasts)
- You can write down your thesis at least 3 months before news coverage
6. What's Next¶
Finding bottlenecks is top-down. But any bottleneck thesis should pass a multi-perspective sanity check — what do Value / Growth / Macro PMs each think?
→ P3-C4 · Multi-PM Perspectives (Value / Growth / Macro) — uses public interviews from 3 types of PMs + a real NVDA case study.
7. Deep Dive (optional): 5 Free Public Sources for Finding Bottlenecks¶
Click to open source map
Company IR Reports + Transcripts (most ground truth): - SEC EDGAR (sec.gov) — 10-K / 10-Q PDFs, all free - Motley Fool transcripts — free 1–2 days after earnings - Company IR websites — Investor Day decks (tons of data) → Search keywords "supply constrained" / "long lead time" / "demand exceeds supply" / "capacity tight"
Sell-Side Reports (partially free): - Goldman / Morgan Stanley / Bernstein occasionally free (leaked on social media) - Free articles on non-paid substacks
Industry Organizations + National Statistics: - SEMI (semi.org) — semiconductor equipment / materials data - U.S. Energy Information Administration (eia.gov) — power data - World Semiconductor Trade Statistics (WSTS) — industry demand data
Free Podcasts: - Acquired (deep industry, occasionally covers semiconductors) - The Information's Tech Briefing (free newsletter)
13F / Holdings Data: - Whalewisdom (free tier, see top fund holdings changes) - Dataroma (free super-investor portfolio)
5-Step Applied to "Power Bottleneck" (Simplified, 2026 Context):
- Identify: 1 GW / data center, U.S. annual new demand 80 GW vs current ~10 GW/year new supply. 8x gap.
- Supply: Nuclear takes 10 years (CEG Three Mile Island announced 2024, online 2028). Gas 3–5 years. Solar + storage 2–3 years.
- Demand: MSFT / GOOGL / AMZN / Stargate all signing PPAs, $50B+ already committed.
- Price: Industrial electricity prices up 8% in 2024 (vs 2023). But power markets are fragmented, unlike HBM transparency.
- Sentiment: CEG / VST / TLN already up 100–300% in 2024. Step 5 signal already fired, your leading edge is gone.
→ Conclusion: The power bottleneck thesis is now too late. Find the next bottleneck (CPO / HBM4 / 2nm yield).