Three Moats Left. The List Is Missing a Fourth.
Nikita Bier compressed a year of founder anxiety into three bullet points this week:
Technical moats are mostly over at this point. If you're building something that is meant to be durable, the only moats now are:
• Proprietary data
• Network effects
• A physical component
I think the first sentence is right, the list is half right, and the most important line is missing from it. In June I argued that the model is not your moat. This is the next step: almost nothing you build on top of the model is either. Here is how I would redraw the list.
A moat is what the copycat still lacks
"Is this hard to build?" used to be the test. It no longer tells you anything, because building is the part that got cheap. The test I use now: a funded team with agents sets out to copy you, and a quarter later, what do they have?
Run each candidate through that test and the list sorts itself into three groups: gone, leaking, and holding.
| Moat | The pressure on it | Verdict |
|---|---|---|
| Technical | Features are rebuilt in weeks. Deep systems last longer, but The Register reported in July on AMD using AI-written code in ROCm.AI to break the CUDA lock-in. | Going, slowest at the deep end |
| Proprietary data | Simulation, open datasets and a million hours of human video are filling gaps that used to take years of collection. Reuters Breakingviews titled a piece "Data guzzler threatens software's last moat". | Leaking |
| A physical component | Nori's two-armed home robot lists at $1,688, VLAI's wheeled humanoid at about $2,900. Unitree is going public at a $9 billion valuation while its hardware lead is being priced away. | Weak on its own |
| Network effects | A clone of a network is an empty room. | Holds |
| Distribution | Supply of products is exploding. Attention is not. | Holds, and missing from the list |
Technical moats are on a clock, not dead
The usual objection is that nobody will vibe-code a database engine, a chip toolchain or a self-driving stack. That is true this year. But it is an argument about time, not about kind. Depth buys you more quarters before the copy arrives. It does not change the fact that the copy is coming, and CUDA, the standard example of an uncopyable software moat, is already being worked on with AI-written code.
What survives inside the technical layer is not the code. It is everything learned while running it: the rules written after each failure, the checks that catch them now, the structure that has been pruned over a thousand runs. Cursor's $60 billion bet was described as a bet "on the harness, not the model", and I think that is the right reading. The model is swappable and the code is rewritable. The accumulated operating record is neither, and it is really a form of proprietary data.
Data leaks faster than people think
Proprietary data is the strongest line on Bier's list and still weaker than it sounds. Data protects you when it is both exclusive and impossible to approximate. The second condition is failing in one field after another. Robotics is the clearest case: the data gap that was supposed to take years of fleet collection is being narrowed with simulation, open datasets and human video.
The data that keeps its value is the kind produced by your own users doing real work in your own product. Nobody can simulate that, because it depends on having the users.
Physical only counts when something comes back
Hardware is hard to make. But "hard to make" is exactly the kind of moat that just stopped working for software, and robot prices fell hard this year. What the serious robot companies are fighting over is not the machine. It is teleoperation hours, fleet logs, and who has the most units reporting back.
So the third bullet is the first bullet again. A physical component protects you when it produces proprietary data, or sits on a supply chain someone else cannot rent. Otherwise it only raises the cost of copying you, and that cost is falling.
The missing line is distribution
When building costs almost nothing, the number of products grows without limit and the number of hours people can pay attention stays the same. The scarce thing stops being the product and becomes being heard and being trusted.
Brand, an audience, a sales channel you own, a licence from a regulator: these look like different things, but they are the same advantage. The buyer already comes to you before comparing alternatives. A copycat can match your product in a quarter and still launch it to nobody.
This is the one moat I would add to the list, and I would put it first. It is also the only one a solo builder can start on today, with no funding and no factory.
The company with no moat
One case refuses to fit any version of the list. "Lovable has no moat" is one of the most common takes in the Valley, and the company reports a revenue run rate above $500 million, growing sevenfold in a year.
I don't read that as proof that moats are a myth. Speed and moats answer different questions at different ages of a company. Speed is how you win users before anyone reacts. A moat is what stops those users leaving once ten clones exist. Early on, hunting for a moat is a trap. Later, not having one is the trap.
What I take from it
Add the fourth line, and put it first. A way to reach people that you own, then proprietary data, network effects and a physical component.
Ask what comes back. For anything physical, the question is not "is it hard to manufacture" but "what does each unit sold teach me that a cloner will not know".
Keep the records, not just the code. Whatever you build with agents, the code is the part a copycat gets for free. What they don't get is the pile of failures, the rules written after each one, and the checks that now catch them. That record starts small, but it is proprietary data from day one, and it only grows if you write it down.
A moat was never the water. It was the fact that there was one bridge, and you decided who crossed it. Agents can dig the same ditch around any castle now. They can't make anyone walk up the road to yours.
Header image: Jan Jacob Teyler van Hall, "Slot Batestein bij Vianen" (1840), Rijksmuseum, public domain.