AI entrepreneurs today fret that new products can be easily cloned, echoing the old “ChatGPT wrapper” meme.
In the words of Sevak Avakians, Responsible AI Principle at USAA, “AI used to be a competitive advantage, now it’s just table stakes”. With big labs able to replicate any clever model, founders wonder how to build truly defensible businesses. Yet, moats are still possible by using classic strategic powers in new ways. Drawing on Hamilton Helmer’s framework, the Lightcone Podcast listed seven moats adapted for AI. We’ve gone a bit further and listed each, illustrated with a modern startup example.
1. Process Power
A process-power moat arises from building an extremely complex product refined over years. As the Y Combinator Lightcone podcast puts it, “You’ve built a very complicated business that’s really hard for people to replicate. Usually this happens when you’re doing something mission-critical, and it takes years of iteration to get right.”
Why it works: Because competitors can easily copy a simple demo, but replicating a full-scale, iron-clad system is extremely costly. Companies like Stripe or Plaid (in fintech) earned defensibility this way: thousands of APIs and intricate backend logic built up over time. In AI, the same applies. For example, Greenlite AI has built an AI platform that “automates manual, mission-critical work” for banks (KYC/AML compliance, sanctions checks). Its finely-tuned AI models and workflows, built with domain experts and real customer data, create a moat: rivals could write a toy KYC demo in days, but Greenlite’s production system took years of data and iteration, making it prohibitively hard to copy. (In other words, the final “10%” of performance (zero false negatives on fraud checks) took 10×-100× the work of a hackathon prototype.)
2. Cornered Resources
This moat comes from exclusive access to a resource that competitors can’t obtain. Cornered resources must be independently valuable and non-arbitrageable. As Helmer’s framework notes, these are “coveted assets… not easily accessible to competitors”. They might be patents or proprietary data, but in AI the most potent cornered resource is often a special partnership or contract that no one else can replicate.
Why it works: If your startup has something like a government contract, proprietary dataset, or a trained model that others literally cannot use (e.g. due to legal or clearance barriers), rivals can’t just buy or copy it. For instance, Scale AI secured a five-year, $100 million Department of Defense agreement to deliver advanced AI tools and data to warfighters. This kind of DoD partnership entails physical SCIF infrastructure, security clearances, and deep relationships in government, which are resources that take years to build. A competitor without those clearances simply can’t step in. Scale’s DoD contract is a quintessential cornered resource moat: it gives preferential access to sensitive data and prevents others from easily entering the space.
3. Switching Costs
A switching-cost moat traps customers in your product because moving to a competitor is too expensive or disruptive. Once a customer embeds their data, workflows, and integrations deeply in your system, even a slightly better rival is hard to adopt. Switching costs create powerful moats when customers become trapped due to the expense and complexity of finding alternative solutions, even if better options exist.
Why it works: High switching costs make customers loath to change: they’ve invested too much (time, data, training) in your solution. In AI, this often emerges via deep, custom integrations. Many vertical AI startups employ a forward-deployed engineering model: long, on-site pilots where engineers build tailored solutions. Over 6-12 months they hardwire the AI into the client’s core processes. The result is that switching would “risk losing a full year of productivity” and a new vendor would have to rebuild all that custom work. For example, HappyRobot (YC ‘21) built an AI agent specifically for logistics workflows (e.g. automated call/email/WhatsApp agents for freight brokers). The founders literally onsite-integrated the product into DHL’s systems, fine-tuning it for DHL’s exact needs. Because HappyRobot’s AI became embedded in DHL’s operations, moving to a generic AI assistant would be extremely painful. HappyRobot thus enjoys strong retention (its pilot-to-contract conversion has reportedly exceeded 95%).
4. Counterpositioning
A counterpositioning moat comes from adopting a strategy that the incumbent cannot copy without self-sabotage. In Helmer’s terms, you do something an incumbent could do, but doing so would “cannibalize their existing business model”. In practice, this often means targeting customer value in a way that conflicts with the incumbent’s current pricing or feature set.
Why it works: By definition, the big players are anchored to their old models. If you position your startup differently, they can’t match you without hurting their own revenues. For instance, many enterprise SaaS firms charge per-seat fees. A counterpositioning startup in that space might automate tasks and thus reduce headcount needed, which would directly shrink the incumbent’s per-seat revenue if copied. A concrete example is Avoca.ai, an AI call center for home services. Avoca chose to focus only on trades (HVAC, plumbing, etc.) and deeply optimize for those workflows. By tailoring objection-handling and on-call rules for HVAC specifically, Avoca can offer a much higher ROI than a one-size-fits-all voice assistant. A general-purpose provider (or a SaaS CRM) would have to overhaul its entire product and pricing to match Avoca, effectively cannibalizing its own broader market. Avoca’s narrow positioning, and the resulting “superior product for HVAC”, illustrates counterpositioning: it exploits the incumbent’s unwillingness to disrupt their wide but shallow offering.
5. Brand
A brand moat means being so well-known that customers pick you over competitors, even if the products are similar. As the podcast notes, brand translates into consumer preference: “Customers choose your product even with equivalent alternatives… well-known brands maintain position through familiarity”. Building a strong brand takes time and cannot be easily copied, giving an advantage especially for consumer-facing AI apps.
Why it works: A trusted name drives user adoption and word-of-mouth. In AI, the classic case is OpenAI’s ChatGPT. Despite Google’s massive user base and resources, ChatGPT quickly became the dominant consumer AI brand. Curiously, even though Google’s Gemini models are, by many accounts, on par with ChatGPT (and Google’s brand is far bigger), ChatGPT has more daily users. Early adopters flocked to ChatGPT out of brand buzz and familiarity, cementing OpenAI’s lead. This shows how a startup can outrun a tech giant: build a strong AI “consumer brand” (via PR, community, speed of iteration) and users will choose you even if an alternative is theoretically similar.
6. Network Effects (Network Economy)
Network effects occur when each additional user makes the product more valuable for others. In AI, the dominant effect is data/network effects: more users generate more data, which improves the model, which attracts more users in a virtuous cycle. Traditional examples are social networks or payment rails; for AI tools, the “network” is often the pool of usage data and feedback.
Why it works: Every time a user interacts, the system learns and improves, raising the barrier for copycats. For example, the Cursor IDE (an AI code editor) explicitly uses user data as a moat: every keystroke and mouse click from its community of developers feeds into training its autocomplete model. As the number of developers grows, Cursor’s coding AI becomes noticeably better for everyone. This creates a compound advantage: new startups must acquire as many users (and data) as Cursor to match its performance, which is extremely hard once Cursor has scale. In short, AI products with built-in feedback loops (chat histories, annotations, user corrections) gain a network effect moat from their dataset’s continual growth.
7. Scale Economies
A scale-economy moat comes from massive upfront investments that give cost advantages as you grow. In Helmer’s words, these moats let you deliver services cheaper than any smaller rival once you’re big. In AI, scale economies mainly show up at the model and infrastructure layer: training a cutting-edge foundation model requires enormous capital, and only the biggest players can afford it.
Why it works: Once you’ve spent the billions on GPUs and infrastructure, your per-query or per-user cost is much lower than any startup that hasn’t. For example, a recent wave of AI startups like Exa are deliberately creating scale economies via infrastructure. Exa built a massive web crawling system (a multi-million dollar investment) to serve search results to AI agents. One crawl of the web can serve all Exa’s customers, spreading the cost far thinner than if competitors crawled separately. Exa exemplifies how AI startups can build scale economies through infrastructure investments that create cost advantages and barriers to entry. A new entrant would have to pour comparable resources into crawling and indexing before serving any clients. Similarly, on the model side, firms like OpenAI and Google have a scale moat: they spent huge sums pre-training their LLMs, so they can now serve inferences at a tiny incremental cost. This massive upfront capital creates a bar that only a few can cross.
Each of these moats: process power, cornered resources, switching costs, counterpositioning, brand, network effects, and scale economies, offers a different route to defensibility. In reality, AI founders often start by racing on speed and execution, then layer in these powers as they grow. But for long-term value, building one or more of these moats is key to preventing your AI startup from being commoditized.
