1. Introduction: The Existential Crisis

• OpenAI's 4o generated 700M+ images in one week, threatening Midjourney and others.
• "If foundation models provide similar capabilities, do we need specialized AI applications?"

2. Dual Threats

• Vertical: Foundation models absorb application features.
• Horizontal: Other startups quickly replicate features.

3. Absorption Pattern

• AI startups gain attention with new use cases, then foundation model upgrades catch up.
• Users question: "Why pay for a separate solution?" -- churn accelerates.

4. The 2x2 Risk Matrix

• Zone 1 (High Risk): Horizontal + simple (basic text generation, simple chatbots)
• Zone 2 (Medium-High): Vertical + simple (domain chatbots)
• Zone 3 (Medium): Horizontal + complex (advanced code gen, Perplexity-like search)
• Zone 4 (Low Risk): Vertical + complex (Harvey for law, Hippocratic AI for healthcare)

5. MCP and Changing Moats

• Anthropic's MCP standardizes LLM-external tool interaction.
• "Connection" is no longer a differentiator. The real moat is in the "last mile."

6. Optimal Specialization Zone

• Too broad = absorbed by foundation models. Too narrow = can't scale.
• Start narrow and deep, then expand gradually.

7. D Factor for Applications

• Most AI applications have D closer to 1.0 than founders think.
• ARR $10M, SaaS multiple 10x = $100M; with D=0.9, actual value = $10M.

8. Four Moat Stages

1. Pre-Moat: Aspirational, no real defenses
2. Moat Construction: Paying customers, partial integration
3. Moat Validation: Survives major LLM launches without churn
4. Mature Moat: Brand, deep integration, regulatory barriers

9. Conclusion

• "In vertical markets, only companies that constantly innovate and build deep domain expertise can capture the trillion-dollar opportunity."