YC's 7 Moats for AI Startups: Speed First, Then the Seven Powers

On Y Combinator's Lightcone podcast, YC partners Garry Tan, Harj Taggar, Diana Hu and Jared Friedman answer a question they keep hearing from college students: if any AI agent can be cloned in a weekend, how can an AI startup ever build a lasting business? Their answer runs through Hamilton Helmer's book 7 Powers: The Foundations of Business Strategy (2016), which Sam Altman and the hosts agree has become one of the most important books for founders, and updates each of its seven moats with AI-era examples from YC companies.

This article is our own summary, with short quotes, from the video's automatic transcript. Watch the full conversation above for everything that was said.

The short version

• Speed comes first, and it is not in the book. Quoting Varun Mohan of Windsurf, the hosts say that at the start the only moat a startup has is speed. Cursor ran one-day sprints in 2023 and 2024, a pace no big lab can match.
• Don't use moats to pick an idea. A moat is defensive: with nothing to defend, it is "just like a puddle in a field." Find a person with a painful problem, solve it, and the powers show up as you build.
• Process power is the most common AI moat: an agent finely honed over years for real-world conditions, like Casetext in legal or Greenlite (KYC) and Casca (loan origination) for banks. The hackathon demo is easy; the last 10% of reliability is the moat.
• Cornered resources now look like real customer data and workflows captured by forward deployed engineers, plus government access like Palantir's and Scale AI's.
• Switching costs have a new AI-native form: long, deeply customized pilots (HappyRobot with DHL, Salient with banks) that turn into seven-figure contracts nobody wants to re-run.
• Counter-positioning hits incumbents that charge per seat: the better their AI agents work, the fewer seats their customers need.
• Brand, network effects and scale still apply: ChatGPT beat Google's head start on brand, Cursor's usage data improves its autocomplete, and Exa's web crawl is a fixed cost it can reuse across customers.

What they said, by moat

Why moats matter now

Diana explains that Helmer taught at Stanford and wrote the book with examples from the 2000s internet era (Oracle, Facebook, Netflix). Jared says the book would make more sense if it were called the seven moats, and that while the examples are dated, the framework is timeless. Garry cites Peter Thiel's line that competition is for losers: without a moat, infinite competition pushes margins to zero and the business dies. The fear driving students, Jared says, is the "ChatGPT wrapper" meme: getting crushed when a big lab decides your product is valuable.

Speed

Diana points to Cursor: Michael Truell told the batch the team restarted the clock every day and tried to ship daily. A large company needs product managers, specs and reviews; Google, she notes, took years to get Bard out. The hosts add that Cursor and Windsurf only had to think about moats once they were at scale and facing Claude Code and Codex, and describe startups, borrowing from a talk with Bob McGrew, as forward deployed engineering teams discovering which verticals are worth defending.

1. Process power

Jared defines it as a complicated business that is hard to replicate simply because you built so much of it, like the Toyota assembly line. In AI it is a mission-critical agent: if Greenlite or Casca fail, banks lose millions, so it is "more like a self-driving car" than a demo. Garry gives Plaid, which supports thousands of financial institutions, and says the same logic explains why Stripe, Rippling and Gusto are defensible. The hosts add that the painstaking final stretch of reliability is work many engineers, including those at the big labs, are not excited to do.

2. Cornered resources

Garry starts with pharma patents and FDA approval, then government work: Scale AI and Palantir spent years in Washington and built secure facilities to get embedded. For startups, the cornered resource is the real data and workflows a forward deployed engineer captures by sitting with a customer, then turning them into prompts, evals and eventually data sets to tune models. Owning a model that does the specific work better is the strongest version, but the hosts stress it is only one possible moat, and that context engineering alone can carry a startup for its first two years.

3. Switching costs

Diana uses Oracle and Salesforce: once your records and workflows live there, migrating costs a year of productivity. AI companies like HappyRobot (logistics for DHL) and Salient (voice agents for banks) run six-month to one-year pilots that build custom workflows, then convert to seven-figure contracts. Jared notes that LLMs can also lower old switching costs by mapping data from one schema to another, while creating a new kind through deep customization of the agent's logic. The hosts add that memory is becoming a consumer switching cost.

4. Counter-positioning

Jared defines it as doing something an incumbent cannot copy without cannibalizing its business. SaaS support tools like Zendesk, Intercom and Front charge per seat, so successful AI agents would shrink their revenue; he thinks founder-controlled companies have the best chance of cannibalizing themselves. Garry describes Avoca, which serves HVAC companies: software typically gets about 1% of such a company's spend, while Avoca captures 4 to 10% by doing customer support work. Harj adds second movers who win on a better product, such as Legora against Harvey in legal AI and GigaML against Sierra and Decagon in support, and Speak against Duolingo in language learning.

5. Brand

The hosts note that ChatGPT has more daily consumers than Gemini, even though Google had all the users and, in their view, Gemini 2.5 Pro and Flash are equivalent models. They call it counter-positioning too: Google had an ads business to protect.

6. Network economies

Diana explains that in AI the network effect takes the shape of data: chat history feeds the next model, and Cursor uses free-tier usage to train its tab autocomplete. For vertical AI agents, private enterprise data and evals create the flywheel.

7. Scale economies

Jared says this has played out mostly at the model layer, where training a frontier LLM is very capital intensive. Harj explains Exa, search for AI agents, which crawls a large part of the web once and reuses it for every customer. Two newer YC companies, Channel3 and Orange Slice, run similar plays.

Final advice

Garry closes with speed again. Look for a specific person whose pain is existential, the kind where they might not get promoted or might get fired, and go from zero to one on that first.

What it means for your business

1. Ship before you strategize. Early on, speed is the moat. Measure how fast you go from customer request to working feature.
2. Win the last 10%. A reliable agent in a regulated or mission-critical workflow is much harder to copy than a demo.
3. Sit with customers. Forward deployed work turns their data and workflows into prompts, evals and a cornered resource.
4. Price on work done, not seats. If AI reduces headcount, per-seat pricing works against you, and your incumbent competitors are stuck with it.
5. Being second is fine. Stripe, DoorDash, Legora and Speak entered crowded markets and won on a better product.
6. Read the book, then forget it until you have something to defend. 7 Powers is useful once you have traction, not to choose an idea.

Video details

• Title: The 7 Most Powerful Moats For AI Startups
• Hosts: Garry Tan, Harj Taggar, Diana Hu and Jared Friedman (Y Combinator)
• Published by: Y Combinator on YouTube (Lightcone podcast), October 3, 2025
• Watch: youtube.com/watch?v=bxBzsSsqQAM
• Note: quotes and figures come from the video's automatic transcript and are lightly cleaned. Check the video for exact wording.

Want to put this to work?

We help owners turn AI trends into a plan, a product and a launch. Book a free call or see our services.