TL;DR: A moat is whatever stops competitors from taking your customers in 18 months when they decide to come for you. Most things founders call moats are head starts. The actual moats keep shifting as technology rewires the cost of building.

What it means

“Moat” is the most overused word in startup strategy. Every deck claims one. Almost none have one. The useful question isn’t do you have a moat? — it’s what specific mechanism prevents a well-funded, motivated competitor from replicating your position within 18 months? If you can’t answer that in one sentence, you don’t have a moat. You have a head start, which is a different and much weaker thing.

Hamilton Helmer’s seven-powers framework is the most rigorous answer anyone has produced. Each of his seven Power types — Scale Economies, Network Economies, Counter-Positioning, Switching Costs, Branding, Cornered Resource, Process Power — describes a specific Benefit-plus-Barrier structure. If you can’t map your “moat” to one of these seven, you probably don’t have one.

But the moat landscape is shifting faster than any static framework can capture. The interesting question now is which of the seven still work in a world where capability is commoditizing.

The argument

Before you have a moat, uncertainty IS the moat. Early-stage startups rarely have any of Helmer’s seven powers. What protects them is that nobody knows yet whether the market is real. Established players won’t pour resources into uncertain markets — the expected value doesn’t clear their internal hurdle rates. The startup’s willingness to operate inside ambiguity is itself a form of defense. You don’t have to be defensible if nobody believes you’re worth attacking (startups-and-uncertainty).

Data moats are real but most “data moats” aren’t. The conventional wisdom is that proprietary data creates durable advantage in AI. Half right. The more important dynamic is that AI itself enables “cheating”: you can use LLMs to generate synthetic training data, bootstrap cold-start problems, and compress the data acquisition timeline. The actual moat isn’t the data — it’s the pipeline that continuously generates and refines it, feeding back into the product so the next conversation is better than the last (cheating-is-all-you-need). Static datasets are not moats. Compounding pipelines are.

When dev costs approach zero, code moats collapse. If AI tools make it trivially cheap to build software, then the software itself stops being a moat. The advantage shifts upstream (to distribution, relationships, brand) or downstream (to customer lock-in, workflow integration, accumulated user context). Every startup whose pitch is “we built something hard to build” needs to update its priors (software-dev-costs-moats, distribution).

Social moats are fragile because network effects can reverse. Not all network-effects are equal. Utility networks (payments, communications) get stronger monotonically — they don’t unwind. Social networks don’t have that property. They run on status dynamics, and status flips. When a platform gets too crowded, the high-status users who made it cool leave for somewhere more exclusive, and the network effect that built it unwinds the same way it built (status-as-a-service).

Community as moat only works with the right substrate. “Community is the new moat” is the slogan of the decade and it’s mostly wrong. Community built around a product that doesn’t independently deliver value is just a Discord server with extra steps. Community amplifies existing value; it does not create it from nothing. The exception is when the community itself is the product, which is a very different business model and much harder to pull off than people pretend.

The synthesis

The pattern across these is uncomfortable: moats are moving targets. What constituted durable advantage five years ago (proprietary code, accumulated user data, network scale) may not hold today. The enduring moats are the ones that compound — where your advantage in period N makes your advantage in period N+1 even harder to replicate. Process Power, deeply embedded switching costs, and continuously compounding data pipelines fit this description. Brand and community do too, but only at scale and only with sustained investment.

The honest answer for most startups: you don’t have a moat yet, and that’s fine. Your job during origination is counter-positioning and speed. Your job during takeoff is to build a real moat before the window closes (see power-progression).

The AI moat question, post-ChatGPT

If you needed empirical proof that the moat landscape has flipped, ChatGPT is it. The leaked Google memo argued that open-source LLMs would erode all proprietary AI advantages (google-no-moat) — and on the model dimension, that prediction was correct. Llama, Mistral, DeepSeek, and Claude have all matched or beaten GPT-4 on benchmarks at various points. ChatGPT is still the verb. Why? Because ChatGPT’s actual moat was never the model. It was brand + scale economies + the relationship effects from years of accumulated conversation history that no competitor can replicate without time.

The lesson: when capability commoditizes, distribution wins. AI startups should win where workflow tools are weak and incumbents can’t “just add AI” (ai-startup-vs-incumbent). aggregation-theory suggests the long-run winners will be those who aggregate demand, not those who control supply of models. ChatGPT did this. Almost no one else has yet.

What links here