YC's 7 Moats for AI Startups:
A Field Guide for Founders
Before you have a product, your only moat is speed. After that, seven moats still hold in AI, and some work differently than they did for SaaS. Here is each one in plain terms, with examples, one question to test yourself, and the YC episode at the right minute.
Updated: November 5, 2026 · Source: YC Lightcone, after Hamilton Helmer's 7 Powers · Read: 8 minutes
At a glance
- The idea: Hamilton Helmer's 7 Powers names the seven lasting sources of business advantage. On YC's Lightcone podcast, Garry Tan, Jared Friedman, Diana Hu and Harj Taggar test each one against AI startups.
- Moat zero is speed. Pre-product you have nothing to defend. Ship in day-long sprints while big companies move in weeks and months.
- What changes in AI: old switching costs fall because agents can migrate data, new ones rise as agents become the customer's operating system, and per-seat incumbents cannot follow you to per-task pricing.
- What does not count: “we fine-tuned a model” and features alone. Code is the easiest thing to copy.
- The punchline: do not use moats to talk yourself out of starting. Find a hair-on-fire problem, ship fast, then deliberately deepen one or two moats.
Watch the episode
The 7 Most Powerful Moats For AI Startups, from Y Combinator's Lightcone podcast (45 minutes). Every moat below links to the minute where it is discussed, and plays here.
13 chapters
- 0:00The moat problem
- 1:30The Seven Powers framework
- 4:20When to think about moats
- 8:40Forward deployed engineering
- 10:18Process power
- 14:34Cornered resources
- 19:30Switching costs
- 24:54Counter positioning
- 31:24The workforce displacement reality
- 34:00Brand and speed as moats
- 37:30Network economies
- 41:00Scale economies
- 43:44Final advice
Moat zero: speed
In the beginning you do not need a moat. You need to move faster than anyone else can. Ship features in day-long sprints, not quarters. Big companies move in weeks and months, so velocity is your only early advantage. The episode's example is Cursor shipping daily while large IDE makers work through long product-spec cycles.
Speed is not a moat in Helmer's sense: it buys a lead, it does not stop anyone following. Its job is to get you to real customers, where the other seven can start. When to think about moats 4:20
The seven moats
In Helmer's framework, each power needs two things: a benefit, which improves your cash flow, and a barrier, which stops competitors copying it. The AI version of each, as discussed in the episode:
A complex, production-grade system that is painful to copy.
- The last 10% to reach 99% reliability takes 10 to 100 times the effort of a hackathon demo.
- That 10% is the boring work: evals, monitoring, recovery flows, compliance and edge cases.
- A competitor can clone your UI in a weekend. They cannot clone years of edge-case engineering and operations.
Examples: CaseText in legal AI, Greenlight in KYC, Plaid with thousands of financial integrations.
Ask yourself: If a strong team copied our demo tomorrow, how many months until they matched us on reliability for our hardest customer?
Access to hard-to-get assets that others cannot easily replicate.
- Regulatory or institutional access: clearances, certifications and government contracts.
- Deep workflow knowledge: forward-deployed teams mapping messy real-world processes.
- Unique data and evals: proprietary data plus the labels for what “good” looks like in one job.
- Not “we fine-tuned a model”. Others use the same base model; they do not have your data, labels or domain intuition.
Examples: Scale AI and Palantir with Department of Defense contracts.
Ask yourself: What do we have that a well-funded competitor could not buy or hire for in the next year?
Moving off you is too painful to be worth it. AI changes which kind of pain counts.
- Traditional (falling): migrating data between CRMs and ERPs. Agents now do the schema mapping and browser automation, so this barrier is shrinking.
- New enterprise (rising): 6 to 12 month pilots build deeply custom workflows. Once you are wired into operations, re-running a bake-off is a huge waste of the customer's time.
- Consumer memory (rising): an assistant that remembers your history, projects and preferences. Switching feels like starting over with a stranger.
Examples: Happy Robot and Salient, wired into their customers' operations.
Ask yourself: If a rival offered the same product for free, what would our customer lose by switching, beyond the data?
Do what incumbents cannot copy without hurting their core business.
- Pricing: legacy SaaS charges per seat. Good AI means fewer seats, so less revenue for them. Price per task or outcome, and you gain from the automation that hurts them.
- Second movers: a later entrant can win by focusing on the application layer and product quality, not on heavy fine-tuning alone.
- It works because incumbents are trapped by their business model and culture. You are not.
Examples: Legora as a later entrant against Harvey in legal AI.
Ask yourself: What would the market leader have to give up to copy us, and would their board let them?
Customers pick you at parity, or even at a slight disadvantage, because you are the default in their mind.
- “ChatGPT” became the verb. That is not a feature; it is a compounding advantage.
- OpenAI owns the “consumer AI assistant” brand even where model quality is comparable, and Google is playing catch-up.
- Brand compounds on top of speed, product quality and clear positioning. It is earned late, not claimed early.
Examples: ChatGPT against Gemini with consumers.
Ask yourself: When a buyer in our category says “we should use one of those”, which company do they name?
More users, more data, better models, prompts and evals, more users. Not a social graph: a feedback loop.
- Every keystroke and accept or reject in a coding tool trains better autocomplete.
- Conversations and ratings feed into the next consumer model.
- In enterprise, usage generates workflow-specific labels and evals that make an agent uniquely good in that environment.
- Competitors can copy your features, not your compounding feedback loops.
Examples: Cursor, ChatGPT, and vertical AI agents inside enterprises.
Ask yourself: Does our product measurably get better for customer 100 because customers 1 to 99 used it? Can we show the curve?
Big up-front fixed investments that give you a lasting lower unit cost.
- Pay the fixed cost once (infrastructure, data, integrations, compliance), then spread it across many customers.
- Frontier model training is the obvious case. Web-scale crawling and infrastructure for agents is another.
- New entrants must spend heavily just to reach your starting line.
Examples: Frontier labs training foundation models; Exa, Channel 3 and Orange Slice building web-scale infrastructure for agents.
Ask yourself: Which of our costs is fixed and large, and does our cost per customer fall as we grow?
When each moat starts to matter
Not every moat is available at every stage. Our reading of the timing:
| Moat | Starts to matter | What a competitor cannot copy |
|---|---|---|
| Speed | Day one | Nothing yet; it is a head start, not a barrier |
| Process power | First production customers | Years of edge cases, evals and ops |
| Cornered resource | Can start before launch | Access, data and labels others cannot buy |
| Switching costs | After the first long pilot | Custom workflows, integrations, memory |
| Counter-positioning | At launch, via pricing | The incumbent's own business model |
| Brand | Late, after you win a category | Being the default name in the buyer's head |
| Network effects | Once usage feeds the product | Your volume of feedback and labels |
| Scale economies | Late, with heavy capital | The fixed cost a new entrant must match |
The YC punchline
Do not use moats to talk yourself out of starting. Pre-product, you have nothing to defend. Find a hair-on-fire pain, ship fast, then pick one or two moats on purpose and deepen them. Trying to build all seven at once is how a seed-stage company ends up with none. Final advice 43:44
Our take
The most useful shift here is in switching costs. AI makes the old lock-in, data trapped in a system of record, easier to escape, and makes the new lock-in, an agent wired into how a team works, much stronger. We argue the same in The Compound Startup Sandwich: own the data below SaaS and the agents above it, and the product surface in the middle stops being the moat.
Process power and AI-style network effects come from the same place: evals and labels for what “good” looks like in one workflow. If you build one eval set this quarter, build it around your hardest customer's edge cases. That is the work that turns into your moat. For the market side, see Build the Environments, Not the Workforce; for ideas YC wants funded, see YC's Request for Startups.
Sources: Y Combinator's Lightcone episode The 7 Most Powerful Moats For AI Startups (October 2025); Hamilton Helmer, 7 Powers: The Foundations of Business Strategy; and Guillermo Flor's founder-friendly summary for Product Market Fit, which this guide's structure follows. Company examples are the ones cited there. The “Ask yourself” questions and the timing table are our own. This guide is independent and not affiliated with Y Combinator. Spot something out of date? Tell us.