Notes: The 7 Most Powerful Moats For AI Startups

This post summarizes a Y Combinator discussion on the most powerful moats for AI startups, based on Hamilton Helmer’s book “7 Powers: The Foundations of Business Strategy.”

The Seven Powers Framework

Hamilton Helmer’s book identifies seven types of competitive advantages (moats) that make businesses defensible. The YC partners discuss how these apply specifically to AI startups, plus one additional moat that’s particularly relevant for early-stage companies.

1. Speed (The Startup’s Original Moat)

While not in Helmer’s book, speed is the foundational moat for startups.

Key insight: Startups can move faster than large incumbents because they have less to lose and fewer organizational constraints.

Example - ChatGPT: OpenAI shipped ChatGPT very quickly with a small team of engineers. Despite Google having all the users, the best AI researchers, and massive resources, OpenAI built the dominant consumer AI brand through speed. If someone had predicted in 2022 that ChatGPT would have more daily users than Google’s Gemini, it would have seemed incredulous.

2. Process Power

Building complicated systems that are hard to replicate, developed through years of iteration.

Classic example: Toyota’s assembly line - competitors couldn’t just copy it because it was built through decades of refinement.

AI application: AI agents that have been finely honed over years of iteration. The more you work with customers and refine your workflows, the harder it becomes for competitors to replicate.

Example - Salient: A company doing AI phone calls for auto dealerships. They’ve spent years perfecting their system, making it extremely difficult for new entrants to match their quality.

3. Cornered Resource

Coveted assets that aren’t arbitrageable - things competitors can’t easily acquire.

Types of cornered resources in AI:

• Patents and regulatory approval - Especially important in regulated industries like healthcare
• Proprietary data - Data that only you have access to
• Custom models - Fine-tuned models trained on unique datasets

Example - Happy Robot: A company doing AI phone calls for trucking logistics. They have access to proprietary data from trucking companies that competitors can’t easily obtain.

Example - Speak: The language learning app has built custom voice models specifically for language education that would be very difficult to replicate.

4. Switching Costs

When customers are “trapped” because switching to a competitor is expensive or painful.

Classic examples: Oracle databases, Salesforce - once you’ve built your workflows around them, migration is extremely costly.

AI application:

• Custom workflows built on top of AI systems
• Data migration challenges
• Integration with existing business processes

Key insight: The more deeply integrated your AI solution becomes with a customer’s operations, the higher the switching costs.

5. Counter-Positioning

Doing something that’s difficult for incumbents to copy because it would cannibalize their existing business.

Classic example: Per-seat pricing vs. work-delivered pricing. If an incumbent charges per seat, they can’t easily switch to outcome-based pricing without destroying their revenue model.

AI examples:

Legora vs. Harvey (Legal AI): Harvey was the early winner but focused heavily on fine-tuning as their differentiation. Legora counter-positioned by focusing on the application layer and building a better product. This worked because Harvey couldn’t easily abandon their fine-tuning approach.

Speak vs. Duolingo: Duolingo is essentially a gaming app with language learning elements. Speak counter-positioned by using LLMs and voice to actually help users practice and learn languages through speaking. Duolingo can’t easily copy this because it would undermine their gamification-based engagement model.

Google vs. OpenAI: Google had a business model requiring them to support ads. OpenAI could build a pure AI assistant without that constraint. Google had “the greatest cash cow in the history of man” - why would they disrupt it?

6. Network Effects

The value of the product increases as more users or customers use it.

Classic examples:

• Facebook - more valuable as more friends join
• Visa - more valuable as more merchants accept it

AI application - Data Network Effects: In AI, network effects often manifest through data:

• More users generate more data
• More data improves models
• Better models attract more users

Example - Cursor: They collect data from every mouse click and keystroke (with user consent in the free version). The more developers use Cursor, the better their autocomplete becomes, creating a compounding advantage.

Example - ChatGPT: All the history from ChatGPT 1, 2, 3, 4, 5 feeds into training the next version, creating a data flywheel.

Evals as Network Effects: For AI startups, evaluations (evals) create a flywheel - you learn what workflows work or don’t work, iterate on your context engineering, and improve. This only compounds with more usage.

7. Scale Economies

Investing heavily to build something big, resulting in lower per-unit costs than competitors.

Classic examples: UPS, FedEx, Amazon delivery network - massive physical infrastructure enables lower costs per delivery.

AI application: This moat primarily applies at the model layer, not the application layer:

• Training state-of-the-art LLMs is extremely capital intensive
• Only a few companies can afford to do it
• Once trained, inference can be offered very inexpensively

DeepSeek’s impact: The DeepSeek announcement was earthshattering because it suggested training frontier LLMs might be cheaper than thought, potentially diminishing this moat. However, they still built on top of large foundation models - the RL part is cheaper, but you still need the expensive base model.

Example - Exa: A search engine for AI agents. Building a web index is extremely expensive, creating economies of scale that make it hard for new entrants.

8. Branding

Becoming so well-known that consumers choose you even when equivalent products exist.

Classic example: Coca-Cola

AI application: Brand is harder to apply directly to startups because it takes time to build. However, its effects are visible:

Example - OpenAI/ChatGPT: Despite Google having equivalent models (Gemini Pro 2.5, Gemini Flash 2.5) and being one of the biggest consumer brands on the planet, OpenAI built the dominant AI brand. ChatGPT has more daily users than Gemini, which would have seemed impossible in 2022.

Key Takeaways for AI Founders

1. Speed is your first moat - Move fast before incumbents can react
2. Build process power through iteration - Years of refinement create defensibility
3. Seek cornered resources - Proprietary data and custom models are valuable
4. Create switching costs - Deep integration makes customers sticky
5. Counter-position against incumbents - Do things they can’t easily copy
6. Build data flywheels - More users should make your product better
7. Consider scale economics - Some businesses require significant capital investment
8. Brand takes time - But can become extremely powerful

The Vertical AI SaaS Opportunity

The partners note that vertical AI SaaS agents will be at least 10x bigger than traditional SaaS because they tap into a different part of company spend - not the finite software budget, but the much larger labor and services budget.

Example - Giga ML: Competing in customer service against Sierra and Decagon. Their counter-positioning is that their product works better out of the box, enabling faster sales and onboarding. AI agents can do customer support not just as well as humans, but better - they’re fluent in 200 languages and infinitely patient.