Moats in the Age of AI
Visit lobstercap.com for more.
If you want to dive deeper and learn more about the new YC F25 batch, join us at our upcoming webinar later today:
Startup founders are more anxious about defensibility now than at any point in the last decade.
The question comes up in every pitch meeting: “What’s your moat?” Founders know the models are commoditizing, they know OpenAI is moving up the stack, and they know that if they can build something in six months, so can ten other teams.
This anxiety is new.
Pre-AI, moats were something you thought about after product-market fit. Now founders obsess about defensibility before they’ve shipped v1, which is mostly a waste of energy that should be spent on customers.
Y Combinator recently put out a framework based on Hamilton Helmer’s “Seven Powers” to help founders think through this. The book is dated but the principles hold.
The YC podcast applies these principles to AI startups and we wanted to take that framework and overlay what we’re seeing in our portfolio. Especially because the gap between theory and practice matters more in AI than it ever has before.
The Seven Powers Framework
Here’s what the framework covers:
Process Power: Building complex, optimized systems that take years to replicate. In AI, this means getting agents to 99% reliability under real-world conditions, which takes 10-100x more effort than a demo.
Cornered Resource: Access to hard-to-acquire assets like regulatory approvals, unique data, or deep customer relationships. The “diamond mine in your customer’s head” counts as much as any physical resource.
Switching Costs: The burden customers face when migrating to a competitor. In AI, this increasingly means deep workflow customization that makes switching feel like starting over.
Counter-Positioning: Doing what incumbents can’t without cannibalizing themselves. SaaS companies charge per seat, AI companies can charge per outcome. That pricing gap alone creates defensive space.
Brand: First-mover advantage that compounds through network effects. ChatGPT has more daily users than Google Gemini despite equivalent models because they moved first and built the default choice.
Network Effects: In AI, this mostly means data flywheels. More users generate more data, which trains better models, which attracts more users.
Scale Economies: The capital intensity of training frontier models creates barriers to entry. Less relevant for application-layer startups building on top of foundation models.
That’s the framework... But what about in practice.
Speed Is the Only Early Moat That Matters
You should not be thinking about moats on day one.
Find a problem, go solve it, work with customers, and build the product. You will stumble upon defensible advantages along the way. Using the moat framework as a reason not to start is paralysis by analysis.
In the early stages of an AI startup, speed is the only moat worth having. Cursor ships features in one-day sprints while big companies need weeks or months. All of these AI startups are essentially forward-deployed engineering teams for the labs, exploring green field territory to figure out what valuable products to build. Once you find something, then you shift to defending and scaling.
The market has compressed in ways that make speed even more critical.
As our piece earlier this week covered, a typical seed round now happens at valuations and revenue milestones that used to define Series A companies.
This means less time to build defensibility through traction alone, and more pressure to move fast while the territory is still open.
The Other Moats That Compound
Once you’ve found something worth defending, three moats matter more than the others:
process power
switching costs
proprietary data access.
Process power is the painstaking work of getting AI agents to 99% reliability in real-world conditions. The last 10% takes 10 to 100 times the effort of a hackathon demo. Case Text in legal AI, Greenlight’s KYC for banks, Plaid’s thousands of financial integrations are not products you spin up over a weekend, and competitors won’t do the edge case work required to catch up.
The AI version of this is a complicated agent that’s been finely honed over years to work reliably under conditions that break everyone else’s system.
Switching costs in AI come in two forms. Traditional switching costs still exist, though AI can actually reduce data migration pain in some cases. The new version is deep customization through six to twelve month pilots that build custom workflows integrating into operations. Once an AI agent is woven into your workflows, switching feels like starting over. For consumer products, ChatGPT’s memory creates the same effect at the individual level.
Proprietary data access remains one of the strongest moats because foundation models can handle public data but can’t replicate decades of integrations, transactional data, or edge-device signals. Forward-deployed engineers mapping customer workflows create datasets nobody else can access. The shape of network effects in AI is really just data accumulation that improves the product for everyone who uses it.
Counter-Positioning and the OpenAI Question
Many investors have asked founders…. So, what if OpenAI builds this?
It’s valid anxiety, especially after AgentKit’s release sent shockwaves through the AI agent ecosystem.
The answer is that counter-positioning still works, but only if you own the workflow or own the data. SaaS companies charge per seat, but AI success means fewer seats and less revenue.
Startups can charge per task or per outcome without cannibalizing an existing business model and this pricing advantage alone creates space that incumbents can’t enter without destroying themselves.
One assumption that is holding true so far is that startups can win by building on top of OpenAI and embedding into workflows and datasets that OpenAI can’t reach. They will have their own customer relationships, their own proprietary data, their own reasons to exist beyond “we have a better model.”
Put simply… If you’re building a chat window fine-tuned by prompt engineers, you’re exposed, but if you’re building systems of record and engagement where users live all day, you’re defensible.
What This Means in Practice
Moats are not theoretical constructs you plan in advance. They’re measurable advantages that show up in the data as your progress:
Network effects reveal themselves in retention and engagement curves.
Switching costs reveal themselves in churn rates.
Process power reveals itself in customer expansion and competitive win rates.
The companies defining the next era of software will be fast to market, AND then build defensible moats.
Speed gets you to the starting line. Everything else keeps you in the race.
Visit lobstercap.com for more.
Today’s newsletter is brought to you by Stable (YC W20).
Most founders don’t think twice about the address they list when they incorporate — until they realize it’s their home address showing up on public records.
Stable fixes that.
It gives you a real business address you can use everywhere: incorporation, IRS, banks, vendors.
It keeps your personal information private and helps your business look professional from day one.
Every piece of mail that comes to that address, Stable digitizes — with AI summaries that tell you exactly what each item is and what needs your attention, all from one secure dashboard.
Want to never think about address or mail management again? Get an address here
P.S.
Get 50% off a Grow or Scale plan for 3 months with code:LOBSTER50.