# The Only 2 Moats That Actually Work in the AI Era
Speed has collapsed the traditional startup playbook. Every AI model trains on similar data. Every engineer can deploy the same open-source frameworks. Every founder can hire from the same talent pool. This convergence means the old moats—technology, talent, capital—no longer stick.
SC Moatti, a partner at Mighty Capital, makes a direct claim in Crunchbase News: only two moat types survive the AI era. Everything else breaks.
The first moat is counter-positioning. This is the old Judo move in business. You deliberately choose a different path from incumbents, not because you lack resources, but because your strategy requires it. Stripe built counter-positioning by targeting developers first, not finance teams. ChatGPT launched free and reached 100 million users before OpenAI monetized, flipping the SaaS playbook. The incumbent can't easily copy this approach without cannibalizing their existing business model.
Traditional tech companies face a bind here. Microsoft can't launch a cheap, ad-supported search engine without undermining Bing's premium positioning. Amazon can't build a bare-bones shipping service without undercutting AWS margins. Incumbents choose to protect their legacy revenue. Startups choose the undefended path.
The second moat is network effects. This one endures because it compounds. Slack's value grows as more teams join. Figma's design platform becomes stickier as more collaborators log in. A16z has written extensively on this, but the mechanism remains simple: the product becomes harder to leave as the user base expands.
AI models themselves lack network effects by default. A better GPT model doesn't get better because more people use it. Moatti's argument cuts against the prevailing narrative that bigger datasets and more compute power create defensibility. They create performance advantages. They don't create moats.
This distinction matters for venture capital allocation. Funds betting on AI startups often chase companies with slightly better models or marginally lower latency. Those advantages evaporate in months. OpenAI releases GPT-5. Anthropic ships Claude updates. Google launches Gemini upgrades. The performance delta closes.
But a startup with genuine counter-positioning or true network effects survives the commoditization cycle. Consider Perplexity AI. The model itself isn't unique. The counter-position is the interface and search-first distribution. Users choose Perplexity not for superior intelligence but for a different way to interact with AI. That's defensible.
The landscape is littered with dead AI startups that had better models but no moat. They raised Series A rounds on benchmarks. They died because benchmarks aren't businesses.
This framework reshapes how founders should think about AI companies. Building "a better model" is not a strategy. Building a different go-to-market, embedding your product into a user workflow so deeply that switching costs spike, or creating a platform where others build and attract more builders, those are strategies.
For investors, the implication is brutal: demand clarity on which moat the founder is building. If the answer is "our AI is smarter," keep walking. If the answer is "only large enterprises can adopt this workflow through our company" or "users will never leave because we've become the social layer for their industry," then you're looking at something real.
The AI commoditization cycle has already begun. What separates winners from the graveyard is not smarter silicon. It's smarter business design.