Skip to content

P1-C7 · Business Model + Value Capture

Core One-Liner

Each link in the value chain has different profit margins — the link with the strongest moat captures the most value. NVDA's 75% gross margin vs AMD's 12% — the gap comes from moat, not product.

AI Industry Knowledge — History → Technology → Value Chain → Business → Applications → Geopolitics

P1-C7 (Part 1, Chapter 7). After this chapter, you'll be able to use a 5-dimension scoring system to judge any AI stock's value capture ability, no longer fooled by "all AI picks-and-shovels are rising."

1. The Problem: Same AI Sell, Why 6x Profit Margin Gap?

Company Gross Margin Operating Margin
NVDA 75% 60%
TSM 55% 47%
ASML 50% 32%
MSFT 70% 45%
AMD 50% 12%
SMCI 15% 9%
CRWV (Neocloud) 30% (Loss-making)

💡 Click ticker → Multi-Source Profile to see full value chain coordinates + upstream/downstream + data.

Both in the AI industry, **AMD and NVDA are both GPU makers, yet profit margins differ by 5x. **SMCI / CRWV are downstream in the value chain, with the lowest gross margins.

→ You can't use "in the AI industry" as a thesis. You must examine value capture ability at each link.

2. The Solution: 5-Dimension Scoring Framework

Dimension What It Asks High Score = King
Gross Margin How strong is your pricing power 70%+ = Extremely strong (ASML / NVDA)
ROIC (Return on Invested Capital) How much does each dollar invested return 30%+ = Sustained high ROIC = real moat
Moat Source Why can't others catch up Physical bottleneck > Ecosystem lock-in > Scale > Brand > No moat
Switching Cost How high is the cost for customers to switch EUV (no substitute) > CUDA (rewrite code) > Server assembly (easily swapped)
Customer Stickiness How likely are customers to leave Platform user stickiness > Long-term contracts > Short-term contracts > Spot market

5-dimension high score → True King (NVDA / ASML / TSM). 3-4 dimension high score → Second-tier (MSFT / GOOGL / Hyperscalers). 1-2 dimension high score → Pick-and-shovel OEM (SMCI / Foxconn — low margin, high volume).

3. How It Works: 5-Dimension Scoring Applied to 6 Role Kings

3.1 ASML — Ultimate Moat (5/5)

  • Gross margin 50%, Operating margin 32% (not exceptionally high for equipment companies, but ROIC 30%+)
  • Physical bottleneck: Only company that can make EUV lithography machines, 25 years of R&D. Canon / Nikon chased for 15 years and couldn't catch up.
  • Switching cost: No substitute. TSM / Samsung / Intel all must buy.
  • Customer stickiness: One EUV machine costs $300M, customers use it for 10 years.
  • 5-dimension score: ⭐⭐⭐⭐⭐

→ True long-term holding. But growth is slow (only 50-100 EUV machines shipped per year).

3.2 NVDA — Ecosystem + Platform (5/5)

  • Gross margin 75%, Operating margin 60% (jaw-dropping)
  • Moat source: 20-year CUDA ecosystem (explained in C3) + Mellanox networking + strategic partnerships + scaling laws positioning
  • Switching cost: Customers switching to AMD MI300 = rewriting all ML code + unfamiliar ROCm + performance gap
  • Customer stickiness: All 4 hyperscalers use it, long capex cycles lock in demand
  • 5-dimension score: ⭐⭐⭐⭐⭐

→ Strongest value capture in the AI era. But biggest risk: Customer in-house ASICs (TPU / Trainium / Maia) divert demand, diluting moat over the long term (10 years).

3.3 TSM — Oligopoly (⅘)

  • Gross margin 55%, Operating margin 47%
  • Moat: Only company with stable mass production at 3nm/2nm globally (Samsung lags, Intel 18A still ramping)
  • Switching cost: High (tape-out + qualification takes up to 2 years)
  • Customers: Apple / NVDA / AMD / Qualcomm all rely on TSM
  • 5-dimension score: ⭐⭐⭐⭐ (minus 1: high customer concentration, AAPL alone accounts for 25%)

→ Rides NVDA's wave. But biggest risk: Taiwan Strait geopolitics.

3.4 MSFT — Hyperscaler + Platform (⅘)

  • Gross margin 70%, Operating margin 45%
  • Moat: Azure scale + Office 365 + GitHub + OpenAI strategic partnership (exclusive until 2024)
  • Switching cost: Enterprise cloud migration takes 1-2 years, difficult
  • Customer stickiness: High (enterprise long-term contracts)
  • 5-dimension score: ⭐⭐⭐⭐ (minus 1: capex black hole, FCF narrowing)

→ Core holding. But capex/revenue ratio 35%+ is a new RISK.

3.5 OpenAI — Users + Brand (⅗)

  • Estimated gross margin ~50% (API 50%+, ChatGPT Plus ~30%), operating deeply loss-making ($5B+/yr burn)
  • Moat: Brand + user base + data flywheel
  • Switching cost: API users switching to Anthropic / Google = 1 week
  • Customer stickiness: Medium (API B2B is sticky, C2C ChatGPT vs Claude / Perplexity is easy to switch)
  • 5-dimension score: ⭐⭐⭐ (minus 2: not profitable + low switching cost)

→ Not a public stock (OpenAI current valuation $852B post-money, per OpenAI official 2026/03/31 announcement, $122B committed capital — OpenAI / Bloomberg; the earlier ~$300B 2025 round has been substantially re-rated, R6 Fix 3). But you must think clearly: OpenAI's moat is weaker than NVDA's, its $852B valuation heavily depends on scaling laws continuing + application layer delivery.

3.6 SMCI / Foxconn — Assembly OEM (⅕)

  • Gross margin 15%, Operating margin 9%
  • Moat: Almost none (server assembly has low barriers)
  • Switching cost: Extremely low (customers can switch assemblers in 1 quarter)
  • Customer stickiness: Low (spot market)
  • 5-dimension score: ⭐ (short-term rides NVDA's wave, long-term commodity)

→ Rises with the tide, falls with it too. SMCI already dropped 60% in 2024 (Hindenburg report + accounting scandal). These are high-beta stocks, not long-term holdings.

4. vs C6 — What You Already Know

Dimension C6 Gives You C7 Adds
Value chain map ✓ Doesn't score who's strong or weak
Value capture ✗ 5-dimension scoring → True kings vs. followers
Investment implication Knows who moves when AI rises Knows who to hold long-term, who to trade short-term

C6 = flat map. C7 = value distribution heatmap. Without C7, you'd buy SMCI as if it's in the same tier as NVDA → long-term pain.

5. Try It: 5-Dimension Score for 3 Stocks

Task (20 minutes): Score each of the 3 stocks below on each dimension (1-5), using knowledge from C1-C6:

Stock Gross Margin ROIC Moat Switching Cost Customer Stickiness
AVGO (Broadcom networking + custom ASIC) ? ? ? ? ?
CRWV (Neocloud) ? ? ? ? ?
CEG (Nuclear power) ? ? ? ? ?

Hints: - AVGO: Multi-business diversified + Google TPU manufacturing + networking — similar to NVDA moat? - CRWV: High customer concentration + Neocloud business model — similar to OAI or MSFT? - CEG: Nuclear PPA 20 years — similar to ASML moat or low?

Self-check (3 items checked → proceed to P1-C8):

  • You can defend your 3 scores with a friend
  • You can explain why NVDA's 75% gross margin is unsustainable at 90% (anti-thesis preparation)
  • You can predict from the 5 dimensions the 5 most stable AI holds for the next 10 years

6. What's Next

You now know who has a moat. But a moat that doesn't deliver ROI can still evaporate (history's 4 AI winters prove this).

Does the application layer have revenue to support $725B/yr in capex?

→ P1-C8 · What AI Applications Look Like Today — see the 4 major application layers + revenue reality.

7. Deep Dive (optional): ASML Moat Deconstruction / OpenAI Cash Burn / Hyperscaler Capex Black Hole Risk

Click to open 3 deep-dive cases

**ASML Moat Deconstruction: EUV light source uses a 50,000W laser to hit tin droplets, generating 13.5nm wavelength plasma light. This system integrates Zeiss optics (no substitute) + Trumpf laser (no substitute) + ASML's own optical path design. The moat isn't one company, it's the entire Western optical ecosystem** — China's attempt to develop domestic EUV is estimated to take 10+ years.

OpenAI Cash Burn Rate: 2024 estimated revenue: $5B. Cost: $9B+. Net loss: $4B+. Main burn areas: (a) Compute (MSFT Azure) (b) Training next model © Talent ($1M+ salary common). → Depends on continuous funding. The private market valuation is now $852B post-money (2026/03/31, $122B committed capital — OpenAI official; current ~$2B/month revenue ≈ $24B annualized + enterprise >40%); it assumes ChatGPT continues upward → $100B+ revenue (2030 estimate). This is a high-beta assumption, R6 Fix 3 reset.

Hyperscaler Capex Black Hole Risk: Microsoft FY2025 ~$80B AI datacenter investment plan (Brad Smith 2025/01/03 blog — "approximately $80B in fiscal 2025 to build out AI-enabled datacenters"; FY2025 actual additions to PP&E $64.55B per 10-K FY2025; FY2026 capex guide in subsequent earnings). Capex/revenue ratio jumped from 12% in 2019 to 35%+ in 2026. Historical comparison: 1999-2001 Cisco / Lucent / Nortel capex also grew this fast, then the dotcom bubble burst, leading to a multi-year inventory cycle. → Key leading indicator: Is hyperscaler capex guidance being raised? If it flatlines or drops = AI chain de-rate trigger.

8. Further Reading (this chapter — business models + value capture)

All free sources, aligned with P5 0-paid policy

Classic papers / primary sources:

Wikipedia (3-10 min):

Videos / public lectures:

Company IR (quarterly reports + investor day):

Podcasts:

  • Acquired — Berkshire Hathaway — Classic business-model case
  • Acquired — Cloudflare — Software business model + moat

Books (library):

  • Pat Dorsey "The Little Book That Builds Wealth" (2008) — Morningstar moat framework; aligns with this chapter's 5-dimension score
  • Aswath Damodaran "Narrative and Numbers" (2017) — How to translate a business model into a valuation
  • Hamilton Helmer "7 Powers" (2016) — 7 power types (scale economies / network effects / counter-positioning / etc.), extension of C7's 5 dimensions

Pair with this chapter's self-check:

After several Damodaran valuation lectures + "7 Powers" + a few years of Berkshire shareholder letters, you should be able to answer "5-dimension score for 3 stocks" and "why profit margins differ by 6x."