Network Effects & Moats

In the AI race, having the best model today doesn't guarantee winning tomorrow. The companies that will dominate long-term are those building economic moats — sustainable competitive advantages that are hard for competitors to replicate. Let's analyze what those moats look like.

What Are Network Effects?

A network effect exists when a product becomes more valuable as more people use it. The telephone is the classic example — the first telephone was useless, but each additional user made every telephone more valuable.

There are two types:

Direct network effects: The product itself becomes more valuable with more users on the same side. Examples: telephones, social networks, messaging apps.

Indirect network effects: More users on one side attract more participants on another side. Examples: more iPhone users attract more app developers, which attracts more users.

For AI products, indirect network effects are more powerful. More ChatGPT users generate more conversations, which provide more feedback data, which improves the model, which attracts more users.

The AI Data Flywheel

The most important network effect in AI is the data flywheel:

1. Users interact with the model
2. Interactions generate data (preferences, corrections, use patterns)
3. Data is used to fine-tune and improve the model
4. Better model attracts more users
5. Repeat

This creates a positive feedback loop — each cycle reinforces the next. The company with the most users collects the most data, builds the best model, and attracts even more users.

However, the data flywheel has diminishing returns. Going from 1 million to 10 million users provides enormous improvement. Going from 100 million to 110 million users provides much less incremental value. This is a form of diminishing marginal returns — a fundamental concept in economics.

Analyzing Competitive Moats

Use the radar chart below to compare AI companies across key moat dimensions. Select companies to see how they stack up.

Types of Moats in AI

1. Data Moat

Companies with unique, proprietary data have a significant advantage. Google has decades of search data. Meta has billions of social interactions. These datasets are extremely difficult to replicate.

Strength: Very strong for companies with existing data. Hard for startups to build.

2. Switching Costs

Once a developer builds their application on the OpenAI API, switching to a competitor requires rewriting prompts, testing, updating integrations, and retraining users. These switching costs create lock-in.

Enterprise customers face even higher switching costs: compliance reviews, security audits, employee retraining, and workflow redesigns.

3. Brand and Trust

In AI, brand matters because users need to trust the system. OpenAI's "ChatGPT" brand has become nearly synonymous with AI chatbots — similar to how "Google" became a verb for searching.

First-mover advantage in brand building is significant, but not insurmountable. Google, despite entering later, could leverage its existing brand trust.

4. Distribution Advantage

Having existing channels to reach users is a massive advantage:

• Google can embed Gemini in Search, Android, Gmail, Docs — reaching billions instantly
• Microsoft can embed Copilot in Windows, Office, Teams, GitHub
• Apple can integrate AI into every iPhone

Startups like OpenAI and Anthropic must build distribution from scratch, which is why partnerships (OpenAI + Microsoft, Anthropic + Amazon) are so critical.

5. Economies of Scale

As discussed in the compute lesson, larger companies achieve lower per-unit costs. This allows them to offer lower prices or invest more in model improvement — both of which strengthen their competitive position.

Winner-Take-All vs. Winner-Take-Most

Will the AI market have one dominant player or several? Economic theory offers two scenarios:

Winner-take-all happens when:

• Network effects are very strong
• Switching costs are very high
• The product is highly standardized

Winner-take-most (oligopoly) happens when:

• Different customers have different needs
• Switching costs are moderate
• Products are differentiated

The AI market is trending toward winner-take-most. Different models have different strengths (coding, creative writing, analysis), different customers have different requirements (privacy, cost, speed), and switching costs are moderate (API interfaces are similar).

Contestable Markets

Even with moats, AI markets remain contestable — new entrants can challenge incumbents if barriers to entry aren't insurmountable. Open-source models (Llama, Mistral) demonstrate this: they reduce the technology barrier, even if they can't match the data and distribution advantages of incumbents.

A market is contestable when:

• Entry costs aren't prohibitively high
• There are no sunk costs that create exit barriers
• New entrants can access similar technology

The AI market is partially contestable — the technology is accessible (open-source models exist), but the data, compute, and distribution advantages of incumbents are substantial.

Key Takeaways

• Indirect network effects and the data flywheel are the strongest competitive forces in AI
• AI moats include data, switching costs, brand, distribution, and economies of scale
• The AI market is trending toward oligopoly (winner-take-most), not monopoly
• Open-source models keep the market partially contestable
• Distribution advantage (Google, Microsoft, Apple) may ultimately matter more than model quality