Skip to content

30 AI Investment Patterns (Pattern Library)

Core of This Page

Repeated patterns extracted across multiple cases — giving you base rates, not predictions. Part 4 gives you "fish" (5 cases), this gives you "fishing" (30 universal patterns).

How to Use These Patterns

Each pattern contains 4 sections: 1. One-sentence rule 2. Base rate (historical batting average / case realization rate) 3. Historical / Current AI case 4. How to use in your thesis (invalidation_triggers / sizing / monitor signals)

30 Patterns in 6 Categories

  • A · Short-term price vs long-term thesis (#1, #3, #5, #14, #18) — 5
  • B · Industry structure (#2, #4, #6, #7, #10) — 5
  • C · Company selection (#8, #11, #23, #25, #24) — 5
  • D · Long-term risk (#9, #12, #13, #21, #30) — 5
  • E · Investment rhythm (#15, #19, #26, #27, #28) — 5
  • F · Geopolitics + Capital (#16, #17, #20, #22, #29) — 5

1. Jevons Paradox · Efficiency Gains → Total Demand Rises

Rule: When a resource becomes more efficient, its total demand increases (not decreases). Because efficiency↑ → price↓ → more use cases → total volume↑.

Cases: - 19th century British coal steam engine efficiency improved → total coal demand rose - 2025/01 DeepSeek training cost -90% → NVDA short-term -17%, but recovered +60% in 1 month as inference demand surged

How to use in your thesis: - See "AI efficiency breakthrough" news → short-term panic is an opportunity, not a thesis breaker - But prerequisite: application layer is indeed expanding (ChatGPT MAU / Copilot ARR still growing) - Invalidation: If OAI / Anthropic revenue growth rate < 50% YoY simultaneously with an efficiency breakthrough → then Jevons fails, real de-rate

2. Incumbents Miss Paradigm Shifts · Strategic Mindset Gap

Rule: Incumbents don't fail due to technology, but because strategic mindset makes them protect profits > bet on the future. True winners kill existing cash cows to capture new platforms.

Cases: - Intel was CPU king in the 90s/2000s, but missed 4 paradigm shifts from 2006-2014 (GPU + mobile + AI) (Otellini saw iPhone "volume too small, rejected") - NVDA Jensen in 2014 cut gaming card profits to pivot to data centers — now 8x returns - Cisco 1999 → 2002 missed the cloud (stuck on hardware), hasn't recovered 25 years later

How to use in your thesis: - Evaluate incumbents: Look at CEO tenure + willingness to kill cash cows - Short-tenure CEO (4-8 years turnover) + large cash cow business → likely to miss next shift - Application: AAPL (Tim Cook 14 years, slow on AI but has cash) vs Intel (4 CEOs in 8 years, missed)

3. Infrastructure Rises First, Applications Later · capex / revenue lag

Rule: In a true tech revolution, the infrastructure layer (selling shovels) rises 1-3 years before the application layer (gold miners). Capex flies first, Revenue materializes later.

Cases: - Dotcom 1995-1999 Cisco / Lucent / Sun rose 100-1000% — then application layer (Amazon truly delivered) took another 5-10 years to soar - AI 2023-2025 NVDA / TSM / VRT / CEG fly — application layer (OAI / Copilot) real revenue delivery only started 2024-2025 - Industrial Revolution 1820-1840 railroads + steel rose first, factories + logistics delivered 1850+

How to use in your thesis: - Early cycle: Heavy on infrastructure (NVDA / VRT / CEG) - Mid-cycle: Shift to applications (MSFT / vertical AI) - Leading indicator: hyperscaler capex / revenue ratio. Once > 40% for 2 quarters = warning

4. Customer Concentration Blowup · Single Customer >50% Risk

Rule: Any company with a single customer >50% of revenue has long-term blowup risk. Customer's economics or strategy changes → company is uncontrollable.

Cases: - ORCL OCI: OpenAI accounts for 54% of RPO ($300B / $553B) - CRWV (CoreWeave): MSFT accounts for 60%+ of revenue - Historical: Foxconn → AAPL concentration risk; Nokia 1999 → concentrated on European telcos

How to use in your thesis: - Mandatorily check customer concentration when writing a thesis (annual report risk factor section + 10-K) - > 50% = red_flag, add invalidation_trigger: "If large customer discloses order reduction OR large customer's own fundamentals deteriorate" - Valuation premium should be discounted (vs comparable with diversified customers)

5. Export Controls → Local Substitution · Regulatory 6-24 Month Lag

Rule: Any geographic/technology restriction is short-term "obvious loser" compensated by other markets; long-term banned region develops local substitutes forming a new thesis.

Cases: - 2022-2024 US H100 not sold to China → short-term NVDA -7~-10%, but other markets compensated; long-term Huawei Ascend 910C + SMIC 7nm emerged - 1980s US semiconductor export controls → Japan / Korea semiconductor rise - 2018 Trump tariff → Vietnam / Mexico manufacturing rise

How to use in your thesis: - Control announced → short-term panic is an entry (NVDA loses 25% China but other markets compensate) - Simultaneously monitor local substitutes in banned regions for private companies (even if public investment is hard, track them) - Long-term: Market fragmentation (US vs China AI dual track), investment must pick a side

6. Physical Bottlenecks = Micro Differences = Massive Divergence

Rule: When 3-5 oligopolists in an industry make similar products, micro differences like qualification pass rates / yield / strikes can determine stock price divergence of ±100%+.

Cases: - 2024 HBM 3 players (SK Hynix / Micron / Samsung): Due to qualification + strike differences, 1-year stock prices +130% / +60% / -20% - 1990s DRAM cycle: 3-4 oligopolists, cycle peak/trough difference ±200% - 2020s TSM vs Samsung Foundry: Yield gap → customer monopoly for TSM

How to use in your thesis: - Don't buy sector ETFs (e.g., "Memory ETF" includes Samsung as a drag) - Stock-pick — look at each company's qualification status + yield + factory outages - Monitor keywords: "supply constrained" / "qualify ongoing" / "ramp delay"

7. Selling Shovels > Gold Miners · Odds Distribution

Rule: In a real gold rush, most miners go bankrupt, but shovel sellers + jeans sellers truly profit. Because miners bet on "finding gold," shovel sellers bet on "increasing number of miners."

Cases: - 19th century California Gold Rush: 99% of miners went bankrupt, but Levi Strauss (jeans) + railroads + banks truly profited - Dotcom 1995-2002: 99% of .com companies died, but Cisco / Sun (shovel sellers) peaked in 1999 but still outperformed Pets.com - AI 2024-2025: NVDA / VRT / CEG (shovel sellers) up 200%+ ; OAI / Anthropic (miners, private) high valuation but not profitable

How to use in your thesis: - Early cycle, heavy on supporting infrastructure (CEG / VRT / GEV / ETN), not foundation models (OAI private) - Look at any mega capex announcement (Stargate / Saudi sovereign AI / UAE), find previously unpriced supporting links

8. Circle of Competence / Discipline of Non-Participation · Buffett Wisdom

Rule: Not participating in a paradigm you don't understand = wisdom, not a miss. Buffett missed Microsoft / Amazon / NVDA, but didn't break his record.

Cases: - Buffett publicly stated in 2024 "I don't know AI economics 10 years out" — didn't take large positions in NVDA / OAI / MSFT - Buffett entered AAPL in 2016 using 5 steps (circle of competence + moat + management + price + holding) - Berkshire Q1 2025 cash $350B+ — waiting for the next AAPL moment

How to use in your thesis: - Only long-term hold if all 5 steps pass; any 1 step fails = trade, not invest - Cash is not laziness, it's waiting for the next 5-step pass moment - Most AI stocks (NVDA / OAI) fail Step 1 (10-year unpredictability) → shouldn't be long-term core

9. Scaling Laws Data Wall Risk · Long-term Wildcard

Rule: Scaling laws (parameters + data + compute → capability) have held for 5-8 years, but data is finite (total internet text ~40T tokens). Once the data wall is hit, synthetic data or multimodality is needed. Risk: model collapse.

Cases: - GPT-2 → GPT-4: 200x parameters, data from ~10B → ~13T tokens - Chinchilla correction (2022): Data must scale equally with parameters - 2026 estimate: global high-quality text ~40T tokens, GPT-6 needs ~100T → not enough - Synthetic data (Meta / OAI trying) risks model collapse (LLM training on itself produces increasingly worse output)

How to use in your thesis: - Long-term (2027+) NVDA / OAI / Anthropic thesis must include data wall invalidation - Monitor: Are foundation model public benchmarks still improving exponentially, or starting to plateau? - Anti-thesis trigger: GPT-6 (hypothetical) capability not significantly exceeding GPT-5 → first break in scaling laws curve

10. Platform Strategy > Product Strategy · 20-Year Ecosystem Lock

Rule: True winners bet on platforms (developer ecosystem / API / standards), not single products. Platform lock-in is a 20-year moat, product lock-in is 2-3 years.

Cases: - NVDA 2006 CUDA: Gave developers free tools, 20-year ecosystem. AMD 2016 ROCm chasing, still hasn't caught up. - Apple 2008 App Store: Platform locked developers + users + data for iOS - MSFT 1995-2008 Windows API: Let others build on top, locked for 20 years - Counterexample: Intel 90s didn't invest in a platform (no CUDA-class software), only sold CPU hardware → 0 ecosystem in AI era

How to use in your thesis: - Evaluate incumbents + new challengers: Look at whether it has a platform strategy + how old the platform is - Platform 5+ years old + active developers = long-term moat - New challengers (OAI / Anthropic): Are users + API a platform? — Medium

11. Customer Self-Developed ASICs Long-Term Dilute Moat

Rule: When customers grow large, they will self-develop chips to reduce costs + lock supply chain. When 1 hyperscaler accounts for 20%+ of a supplier's revenue, this trend is inevitable.

Cases: - Apple 2008 → A-series chip (manufactured by TSM, not reliant on Qualcomm) - Google 2016 → TPU (in-house, not reliant on NVDA internally ~50%) - AWS 2020 → Trainium2 (replacing some NVDA) - Microsoft 2024 → Maia 100 (internal testing)

How to use in your thesis: - Evaluate NVDA / similar midstream companies: Look at customer concentration × customer chip-making capability - 4 major hyperscalers all doing ASICs → NVDA long-term (10 years) internal market share 60% → 40% - But general market + agentic still NVDA — anti-thesis is not "NVDA dies," it's "growth decelerates"

12. Price Commoditization · Foundation Model Hits a Wall

Rule: Software / API products are easily replicated, price drops of -90% are almost inevitable. Only user experience + data flywheel can lock in.

Cases: - OAI GPT-3.5 API 2022 $0.002/1K token → 2024 nearly $0 - DeepSeek 2025 slashed API price by -90% - AWS S3 2006 $0.15/GB → 2024 $0.023/GB (-85%) - Counterexample: Apple iPhone price not commoditized (brand + ecosystem)

How to use in your thesis: - Foundation model layer (OAI / Anthropic): Long-term thesis is not API revenue, it's application layer monetization (ChatGPT C-side + enterprise SaaS) - High API revenue valuation is unsustainable - Evaluate: User volume + data flywheel (ChatGPT 300M MAU is OAI's true moat, not the API)

13. Energy is a Wildcard but Construction Cycles are Long

Rule: Software can scale quickly, infrastructure (power / real estate / manufacturing) cannot. 1 GW data center takes 3-5 years to build, 1 nuclear plant takes 10 years. Any AI growth ultimately hits physical limits.

Cases: - MSFT-CEG Three Mile Island restart: Announced 2024, online 2028 (4-year lag) - TSM Arizona Fab 21 Phase 1: Announced 2020, production 2024, Phase 2-3 until 2028 (8 years) - AI 2026 US annual new electricity demand ~80 GW vs current annual new ~10 GW → 8x gap - Stargate $500B data center needs 5 years to fully build

How to use in your thesis: - Energy / manufacturing tickers (CEG / GEV / VRT / EQIX) are 4-8 year theses, not 6-month theses - But already priced in early (CEG 2024 +200%) — be cautious adding now - Look for next undiscovered bottleneck (HBM4 2026 / CPO 2026 / liquid cooling penetration rate)

14. 13F Lags 6 Weeks + Doesn't Show Options · Signal Decay

Rule: 13F holdings are disclosed with a 45-day lag (quarter-end → 5/15), and don't show shorts / options. Using 13F as a timing signal → systematic error.

Cases: - 2024 Aschenbrenner $1.57B NVDA put — rare 13F disclosure case (most similar shorts are invisible) - 2026 5/15 discloses Q1 holdings, but market has moved 6 weeks, trigger conditions may have changed - Bridgewater reducing + Citadel increasing same ticker — divergent signal, not a clear buy/sell

How to use in your thesis: - 13F signal: Look at institutional consensus trend (not single fund) - Caveat must be added: "Lags 6 weeks, doesn't show options" - Cross-reference with options open interest (abnormally large strike puts/calls = stronger signal than 13F)

15. Agentic Compute · New Growth Curve

Rule: Regular ChatGPT single session ~1K token inference, agentic (Claude Code / Cursor / Devin) single session ~50K-500K tokens + multi-step tools. Inference compute 100-500x.

Cases: - Claude Code actual inference usage vs regular chat: 100-500x compute - Cursor coding session: average 50K tokens / task - Devin (autonomous agent) each task: hours of LLM calls

How to use in your thesis: - Add new engine to NVDA long-term thesis: agentic penetration rate - Assume agentic penetration 5% of global developers → inference compute current 10x — NVDA second growth curve - Monitor: Agentic company ARR growth rate (Cursor / Anthropic Claude Code revenue)

16. Hyperscaler Capex Inflection Point · Biggest Wildcard

Rule: 90% of the AI industry chain depends on the 4 major hyperscaler capex. Any one company's capex guide flat or down = entire industry chain de-rate 30-50%.

Cases: - Historical comparable: 1999 Cisco capex flat then -90% (dotcom) - Current MSFT/GOOGL/AMZN/META FY26 total capex $290B+, consistently upward - Once FY27 any one company < 10% YoY growth → AI thesis 5th winter triggered

How to use in your thesis: - This is the most important invalidation_trigger (write into all NVDA / hyperscaler chain theses) - Monitor: Quarterly earnings capex guide (4 companies × 4 quarters = 16 data points / year) - Once triggered → trim 50% AI infra position

17. Geopolitics → Market Fragmentation · Dual Track Development

Rule: US-China AI war → within 5-10 years global AI industry splits in two: US + allied market (NVDA / OAI) vs China market (Huawei / DeepSeek). Investors must pick a side.

Cases: - NVDA 2022-2025 lost 25% China market, US friend-shoring filled the gap - DeepSeek trained near GPT-4 level using H800 + algorithm efficiency — proof of China self-sufficiency - Yangtze Memory HBM 2026 mass production → China memory self-sufficiency - Historical comparable: 1980s semiconductor US-Japan split; 2000s+ internet US-China split

How to use in your thesis: - Any thesis involving cross-region exposure must label: "China share X% (could be cut to 0)" - TSM thesis must add Taiwan Strait invalidation (low probability catastrophic) - US friend-shoring beneficiaries: TSM Arizona / Korea / Japan / EU

18. Q4 + Q1 Effect · Year-End/New-Year Catalyst Concentration

Rule: AI / tech industry Q4 earnings concentration + year-end GTC + Q1 NVDA earnings = Dec-Feb historical outperformance.

Base rate: 2014-2024 SOX semiconductor index Dec-Feb average +8.5% (vs full-year average +12%), Dec-Feb contributes 70%+ of annual return.

Cases: - 2024/12-2025/02: NVDA +18% (Q3 FY25 + GTC + fiscal year-end positioning) - 2023/12-2024/02: NVDA +30% (Q3 + Q4 earnings dual catalyst) - Counterexample: 2025/01/27 DeepSeek selloff broke the rhythm

How to use in your thesis: - Q4-Q1 add window, but wait for pullback (not chase up) - Monitor November NVDA Q3 earnings + January GTC + February Q4 = 3 consecutive catalysts - Anti-thesis: If Q3 earnings + GTC both disappoint → weak entire cycle

19. Open Source Wins Long-Term · Platform-Type Moat

Rule: When a layer becomes a pervasive platform (Linux / Android / database / LLM), long-term open source options will cannibalize 50%+ of the paid leader's share.

Base rate: Historical 7/10 cases materialize (slow, 5-10 years). Linux servers 90% / Android smartphones 70% / MySQL+PostgreSQL DB 50%+.

Cases: - 2024-2025 Meta Llama / DeepSeek / Mistral open source → cannibalizing part of OAI/Anthropic API market - Long-term trajectory: foundation model layer commoditization is inevitable - But application layer (ChatGPT C-side / Claude Code agentic) remains closed king

How to use in your thesis: - Foundation model companies (OAI / Anthropic) long-term API revenue thesis is weak - True moat is in application layer + data flywheel + user volume - Monitor: Llama 4 / DeepSeek V4 launch, API prices further compressed → verification

20. CapEx → Depreciation 5-Year Lag · Inflated Profit Trap

Rule: A company's current year capex is typically depreciated over 5-7 years into P&L. High capex companies' current reported profit is inflated — true cost only reflects 5 years later.

Base rate: 1999 Cisco / 2024 hyperscalers both triggered. When capex / revenue > 30% for 3 consecutive quarters, subsequent 5-year EPS growth typically plateaus or turns negative.

Cases: - MSFT FY24 capex $55B vs depreciation $20B → profit inflated by ~$35B (book) - 5 years later when capex is fully depreciated, EPS growth rate assumption -20-30% - Historical Cisco 1999 capex $4B, 2002 depreciation hit → EPS -60%

How to use in your thesis: - Look at hyperscaler actual owner earnings = reported earnings - (capex - depreciation) - MSFT "true" earnings are 35%+ lower than reported - Anti-thesis trigger: When capex / revenue stays > 35% and revenue growth decelerates → re-rate

21. Cyclical Industries 18-24 Month Half-Life

Rule: Memory / semiconductor equipment / some SaaS are cyclical stocks, historical cycle 18-24 months. Peak → trough typically -50% to -70%, then 24-month recovery.

Base rate: Memory cycle (1990-2024) total 8 cycles, average 22 months. Full decline: -60%.

Cases: - 2022 memory cycle: MU peak $99 → trough $48 (-52%), then 24-month recovery to $130 (peak) - After HBM surge in 2024, 2026 H2 may start next memory cycle peak - Semiconductor equipment (ASML / AMAT) cycle syncs with foundry capex

How to use in your thesis: - Cyclical stocks P/E used inversely (low P/E at peak, high P/E at trough) - MU / semiconductor equipment don't hold long-term, treat as cycle trades - Monitor SEMI equipment book-to-bill ratio (>1 ramp / <1 peak)

22. First Mover ≠ Winner

Rule: The first to launch a product is usually not the ultimate winner. The winner is right timing + platform strategy + execution, not first.

Base rate: Historical 6/10 cases, first mover gets overtaken. AltaVista → Google. Friendster → Facebook. Netscape → IE → Chrome. MySpace → Facebook.

Cases: - Transformer 2017 was written by Google, but OAI 2022 captured users with ChatGPT (5-year lag) - Vision Pro (Apple) vs Quest (Meta) — Apple "first premium" but Quest already 6 years + user base - IBM Watson 2011 was LLM pioneer, but 14 years later doesn't own the market

How to use in your thesis: - Don't pay a premium for first mover (e.g., AltaVista 1998 valuation) - Look at who has distribution + capital + execution, not who launched earliest - Counterexample (Winner is first): AAPL iPhone, NVDA CUDA — first and sustained investment

23. Switching Cost > Product Quality

Rule: Once a customer has deep integration + high data migration cost, even if a competitor's product is better, most customers won't switch.

Base rate: 9/10 enterprise SaaS / industrial software cases materialize. Microsoft Office vs Google Workspace 25 years. Adobe Creative Cloud vs Figma 5 years.

Cases: - CUDA (20-year ecosystem): AMD MI300 performance may match H100, but rewriting ML code is too costly - Snowflake / Databricks: Customer data migration 6 months, 90% of customers don't switch - SAP / Oracle DB: 30-year lock-in, even if cloud-native is better

How to use in your thesis: - When evaluating moat, switching cost is the strongest leading indicator - NVDA moat is not the GPU itself, it's the CUDA ecosystem + customer ML code integration - Counterexample warning: AWS S3 / GCP Storage — switching cost moderate, price commoditized

24. Acquisition as Moat Extension

Rule: True winners use acquisitions to extend a single-product moat into an ecosystem moat. NVDA-Mellanox / Adobe-Figma / MSFT-OpenAI investment.

Base rate: 5/10 large acquisitions (>$5B) succeed long-term. Failures often due to culture mismatch. In the AI era, strategic equity (NVDA + CRWV) is more flexible than traditional M&A.

Cases: - **NVDA-Mellanox (2019, $7B): InfiniBand networking → full-stack data center fabric control - **NVDA-Run:ai (2024, $700M): K8s orchestration → locks enterprise AI deployment - **NVDA strategic investments in CRWV $36B / NBIS / Cohere — indirectly locks customer demand - Adobe-Figma** (2022) — $20B (blocked by regulators) — lesson: acquisitions also have risk

How to use in your thesis: - Long-term hold candidates: Look if incumbents use M&A to continuously extend moat - Monitor NVDA Q1 FY27 onwards for new acquisition / strategic investment announcements - Anti-thesis: If NVDA stops M&A → moat growth plateaus

25. Founder-Led Companies Premium 30%+

Rule: Companies long-term led by their founders historically outperform professional managers by 30%+. Due to long-term perspective + risk appetite + internal ownership.

Base rate: Bain 2020 study of founder-led S&P 500 companies vs non-founder, 15-year outperformance +3.1% annualized = cumulative ~50%+. Especially relevant in the AI era.

Cases: - NVDA Jensen 33 years: Cut gaming card profits for data centers + CUDA 20 years + Mellanox acquisition - Meta Zuckerberg 20 years: Reality Labs burning $50B without blinking (professional CEO wouldn't dare) - TSLA Musk: Risk appetite + long-term visionary - Counterexamples: GE / Intel / Boeing — long-term professional CEOs, lagging in AI era

How to use in your thesis: - Founder-led: Prioritize long-term hold (NVDA / META / TSLA) - Look at CEO tenure: > 10 years = long-term perspective ; 4-8 years turnover = short-term KPI mindset - Anti-thesis trigger: Founder departure / health event → re-rate

26. Activist Investor as Catalyst (6-18 Month Timeline)

Rule: Activist entry (Pershing Square / Elliott / Starboard) typically catalyzes specific actions within 6-18 months: spin-off / buyback / CEO change / asset sale.

Base rate: Historical 6/10 cases, target company outperforms market by +15%+ within 18 months of activist entry. Particularly effective in underperforming incumbents.

Cases: - Ackman (Pershing) → CMG 2016 catalyst, 3 years +200% - Elliott → AT&T spin-off WarnerMedia - Starboard → CSCO, pushed buyback + capital return

How to use in your thesis: - Monitor 13F for large activist new positions - AI industry activist candidates: Intel (already happening) / IBM / Oracle non-cloud parts - Anti-thesis: Once activist exits → gains may reverse

27. R&D × ROIC 4-Quadrant Classification

Rule: Companies can be classified into 4 quadrants by R&D / revenue ratio × ROIC. Only "High R&D + High ROIC" are compounding companies.

Base rate: S&P 500 history, high R&D + high ROIC quadrant accounts for 5%, 15-year outperformance +50%+ vs S&P.

Quadrant R&D / Rev ROIC Type AI Example
Compounding (best) High (>10%) High (>20%) Platform + execution NVDA · MSFT · GOOGL
Burning High Low Still burning, validating OAI · Anthropic · Pure Storage
Defensive Low High Mature + cash KO · JNJ · AAPL (post-2020)
Death Spiral Low Low Dying Sears · Boeing · INTC (current)

How to use in your thesis: - Long-term core hold: Only select Compounding quadrant - AI era new compounding candidates: NVDA / MSFT / GOOGL / META - Anti-thesis: If NVDA enters "High R&D + ROIC declining" → leaves compounding quadrant, reduce position

28. Catalyst Clustering · Multiple Events in 1 Week = Positioning Window

Rule: AI industry catalysts often cluster within 1 week (earnings + GTC + policy + 13F). This is an institutional positioning window, retail should reduce position 1 week before + review 1 week after, not add during events.

Base rate: 2023-2025, NVDA earnings week (Q3 + Q4 + Q1) average ±10% volatility, 65% probability of reversal within 7 days post-earnings (short-term buyer fatigue).

Cases: - 2025/02 NVDA Q4 + GTC + DeepSeek event within 1 month: NVDA -17% -> +30% recovery -> then -10% (net -5%, high volatility) - 2024/05 NVDA Q1 earnings + 13F disclosure 1 week later: Even with earnings beat, 13F reduction exposed → sideways - 2023/11 NVDA Q3 + Sam Altman resignation OAI drama within 1 week

How to use in your thesis: - 1 week before catalyst clustering reduce starter to core ratio - Don't move during events, wait 7-14 days for sentiment to settle before reviewing - Monitor catalyst calendar, find next clustering window (e.g., NVDA earnings + GTC + major conference same week)

29. Forward Guide Value > Current Quarter

Rule: Stock reaction on earnings night is 80% driven by guidance, 20% by current quarter. Even with a big current quarter beat, guidance miss → decline.

Base rate: 2018-2024 SP500 post-earnings reaction analysis, guidance miss averages -8%, even when current quarter beats. Conversely, guidance beat averages +6%.

Cases: - NVDA Q4 FY25 (2025/02): Current quarter revenue beat ($39.3B vs $38.5B expected), but Q1 guidance midpoint $43B vs $44B expected → next day -8.5% - ORCL Q1 FY26 (2025): Current quarter miss, but RPO $553B beat + strong AI guidance → +12% - MSFT Q1 FY25: Current quarter beat but Azure growth guide decelerated → -6%

How to use in your thesis: - Your thesis's catalyst look_for should emphasize guidance > current quarter - On earnings night, don't look at headline "beat / miss" totals, look at management guide language - Monitor keywords: "above prior guide" = strong; "in line with prior" = neutral; "moderating" / "headwinds" = warning

30. Black Swan Defense · Sizing Leave 30% Cushion

Rule: Historically, every 7 years on average there is a -30% systemic drawdown (1987 / 2000 / 2008 / 2020). Long-term investing must assume it will happen, leave cushion.

Base rate: 1928-2024 SP500 data, average 7 years 1 -30%, average 10 years 1 -40%. AI concentration high → personal portfolio drawdown larger.

Cases: - 2020/03 COVID -34% in 5 weeks - 2022/01-10 -25% (Fed rate hikes + inflation) - 2025/01/27 DeepSeek selloff NVDA -17% in 1 day - Any event that can take away 1 year of returns

How to use in your thesis: - Portfolio leave 20-30% cash / bond (even if you have high conviction) - Single stock max 15-20% (P2A-C4 sizing already taught) - Mental preparation: Assume tomorrow -30%, can your sizing handle it? - Black Swan happens — don't predict but leave room

30 Patterns' Common Lessons

Lesson Source Patterns
Short-term price ≠ long-term thesis #1 Jevons / #5 regulatory lag / #18 Q4 effect
Look at second-order effects #1 / #11 ASIC dilution / #14 13F lag / #20 capex lag
Incumbent strategic mindset determines paradigm #2 / #10 / #22 first mover
Physical constraints ultimately dominate #6 / #13 energy / #16 / #21 cyclical
Selling shovels > gold mining (1-3 year cycle) #3 / #7
Don't invest in what you don't understand #8 / #9 scaling uncertainty
Compounding = R&D × ROIC × founder-led #25 / #27
Black Swan always happens, leave cushion #30
Long-term platform > single product (open + closed) #10 / #19 / #23
Catalyst timing > Catalyst direction #18 / #28 / #29

Patterns are not predictions, they are base rates

Each pattern gives you historical base rates, not certain future outcomes. You still need to write your own thesis — patterns help you avoid typical pitfalls, not make decisions for you.

Base rate sources: Bain & Co 2020 founder-led study / SP500 historical earnings reaction (Bloomberg consensus 2018-2024) / SOX semiconductor index historical seasonality / Personal observation of AI industry 2022-2026.