What is a competitive moat?

A competitive moat is a durable advantage that makes it hard for rivals to take your customers, the same way a moat around a castle keeps attackers out. Classic moats include a technical lead, a network effect, high switching costs, a strong brand, economies of scale, and regulatory barriers. The test of a real moat is time: it has to keep protecting you after competitors know exactly what you built and try to copy it. Razorpay CEO Harshil Mathur frames it simply, as a barrier to entry that slows everyone who wants into your market.

Is AI destroying competitive moats?

AI is compressing many of them, not all. The moats built on being first to a feature or holding a temporary technical lead are the most exposed, because AI lets a competitor rebuild a working product in days rather than months. Y Combinator founders describe the decade-old moat protecting legacy software as no longer effective for the same reason. What survives are moats that AI cannot copy on its own: customer trust, deep relationships, regulatory approval, and the speed to keep moving before incumbents react.

What kind of moat still works against AI?

Trust is the strongest one. In B2B especially, a buyer is handing you money, data, or a core workflow, and no model regenerates the years of reliable delivery that earns that. Regulated complexity is a second: in payments, health, or finance, every entrant has to clear the same slow, expensive approvals, which thins the field. Speed is the third and most controllable: a small AI-native team can ship and adapt faster than a large incumbent, and staying ahead becomes the moat when static advantages erode.

How do founders build a moat when AI copies everything?

Stop treating the feature as the moat, because that is the part AI copies fastest. Build the advantage in the layer models cannot supply: understand a specific customer better than anyone, earn trust through reliable delivery, and compound both over time. Harshil Mathur's own story shows the pattern, as Razorpay's edge came from early adoption of a new payment rail and years of trust with customers, not from a feature nobody else could build. Choose a problem you will still care about in ten years, then out-execute on speed and depth.

Competitive Moats in the AI Era

A competitive moat used to be the whole point of strategy. You found a defensible advantage, a technical lead, a network effect, a brand, a regulatory wall, and it bought you years while competitors struggled to catch up. AI is draining a lot of those moats fast. A feature that took a team six months to build can be rebuilt by a competitor in a weekend, so the technical edge that once protected you evaporates almost as soon as it is visible.

Harshil Mathur, co-founder and CEO of the Indian fintech Razorpay, made this the center of a Y Combinator conversation titled "AI Is Compressing Every Moat." His view is not that moats are dead. It is that most of the ones founders chase were never as durable as they looked, and AI just exposed it. A few moats do still hold. They are not the ones you build in the product.

What a competitive moat actually is

Strip away the jargon and a moat is a barrier to entry: something that makes it hard, slow, or expensive for a competitor to do what you do. Mathur describes it plainly as the thing that protects a company from its competitors. The classic list is a technical lead, a network effect, high switching costs, brand, economies of scale, and regulation.

The real test is not whether an advantage exists today. It is whether it survives after a competitor knows exactly what you built and sets out to copy it. That is the test AI keeps failing for the product-feature moats. If a capable team with the same models can rebuild your feature in a week, it was never a moat. It was a head start.

Why AI is compressing the classic moats

The moat that AI erodes fastest is the one built on being first to a capability. When building software gets dramatically cheaper, the gap between "we have this feature" and "everyone has this feature" collapses from years to weeks. Mathur's line is that the only way to survive is to figure out what the market is going to be and move today, because the window to hold any single advantage keeps shrinking.

This is not one founder's hunch. Across the founder conversations we track, the same pattern shows up. Y Combinator founders describe the traditional moat that protected legacy software, built up over a decade, as no longer effective in the AI era. And economically, when a capability gets easy to reproduce, its price falls toward the cost of the compute behind it, so the advantage of simply having it does not last. Benedict Evans has made the related point that a lot of the value AI creates flows to customers as surplus rather than staying with any one vendor.

The takeaway for founders is uncomfortable but clarifying: stop treating the feature as the moat. It is the part AI copies first. The defensible part is somewhere else.

Trust is the moat AI cannot copy

Mathur's sharpest point is that B2B is a business of trust, and nothing replaces the human touch point of trust. When a business buyer signs up, they are handing you money, sensitive data, or a workflow their own company depends on. No model regenerates the years of reliable delivery that earn that decision.

Razorpay learned this the hard way. When a banking partner pulled support and threatened to break the product, the company chose to communicate transparently with customers through the crisis rather than go quiet, and that openness is part of what kept the trust intact. That kind of credibility is expensive to build and slow to transfer. A competitor with an identical feature set still starts from zero on it.

Trust compounds, which is what makes it a real moat. In B2B the buyer keeps asking a simple question, whether the service adds more value than it costs, and every reliable month answers it again. This is the same reason the model you build on is a swappable input while your real advantage sits above it: the customer relationship and the understanding behind it are the durable layer, not the technology of the week.

Regulated complexity still slows everyone equally

The second moat that holds is regulation. Mathur points out that in a regulated industry, every competitor has to navigate the same hurdles to get in, so the complexity itself becomes a barrier. Payments, health, lending, and finance all carry approvals, licenses, and compliance that take real time and money to clear.

This one cuts against the AI story in a useful way. AI can help you build the product faster, but it does not shorten a regulatory approval or absorb the legal accountability that comes with a license. The paperwork moves at the speed of the regulator, not the speed of your code. That slowness is annoying when you are the one clearing it, and it is protection once you are through, because it thins the field of everyone who was not willing to.

The caveat is that a regulatory moat protects a market you have already entered. It does nothing for the years of grind to get in, and it can lull you into complacency if you treat clearance as the finish line rather than the entry fee.

Speed is the moat you control

If static advantages erode, the moat becomes motion. Mathur's own example is Razorpay's early adoption of UPI, India's real-time payment rail. By moving on it early, the company captured meaningful market share before larger competitors responded. The advantage was not the rail itself, which was available to everyone. It was being first to build on it and fast to iterate.

He has a name for why incumbents lose these races: the incumbent fallacy, the tendency of established companies to respond to change too slowly and miss the opening. Big companies have process, politics, and legacy systems that make fast moves hard. A small team does not. Founders we follow describe exactly this edge: a startup can design its workflows, tools, and culture around AI from day one and operate far faster than an incumbent bolting AI onto old machinery. That structural speed is a moat a large competitor cannot easily copy, because copying it would mean dismantling how they already work.

Speed is also the most controllable moat on this list. You cannot manufacture ten years of trust overnight, and you cannot skip a regulator's timeline, but you can decide to move first and iterate faster starting today. This is the same reason it pays to go all-in on AI across how you run the company: the operating speed you build now is the defense that holds when the feature-level advantages do not.

Build the moat in the layer AI cannot reach

Put the three together and a pattern appears. The moats that survive AI are the ones a model cannot supply on its own: trust earned over years, regulatory ground already cleared, and the speed of a team wired to move. What they share is that they live in the relationship and the execution, not in the feature list.

That reframes the founder's job. The durable advantage comes from understanding one specific customer better than anyone else and delivering for them reliably enough that they stay. Mathur's test for whether you can build that kind of depth is a good filter: is there a problem you could spend all your time and effort on for the next ten years. Because AI makes the building easy, the scarce thing becomes the willingness to go deep on a problem and the trust you accumulate while you do. That is the moat worth building.

What to do this week

List your top three advantages, then ask of each one: could a capable team with the same AI models copy this in a month? Anything that fails is a head start, not a moat.
Find your trust surface. Identify the one relationship, dataset, or workflow customers would be most nervous to move away from, and invest in making it more reliable, not more feature-rich.
If you are in or near a regulated market, treat the compliance you have cleared as an asset and map what it would cost a competitor to match it.
Pick one place you are moving slower than a startup should and remove the drag this week. Speed is the moat you can build starting now.
Write the ten-year version of your problem statement. If you cannot commit to it, your moat will be shallow no matter what you build.

Moats in the AI era are less about what you build and more about how you operate: what you earn from customers, what you clear, and how fast you move. That operating model is exactly what we teach founders in AI Operating System for Startups.

Sources

Harshil Mathur: AI Is Compressing Every Moat, the Y Combinator conversation this article distills.
Background on Harshil Mathur and Razorpay: Y Combinator's Q&A with the Razorpay co-founders and The Economic Times on Mathur at YC Startup School.
Cross-source on AI eroding moats: Y Combinator, "SaaS Challengers" (the decade-old legacy-software moat no longer holds), Y Combinator, "How To Build A Company With AI From The Ground Up" (AI-native teams operate faster than incumbents), and a16z with Benedict Evans, "The Economics of AI Usage" (much of AI's value flows to customers as surplus).