Building Real Moats for AI Startups

AI startup founders face a real fear: getting crushed by incumbents or dismissed as "ChatGPT wrappers." Foundation models have commoditized basic capabilities, making it harder to find defensibility. Hamilton Helmer's "Seven Powers" framework offers a way through this (https://7powers.com/)—but its application has changed. According to Y Combinator's analysis in "The 7 Most Powerful Moats For AI Startups," here's how these principles translate to modern AI companies. The fundamentals of building a moat haven't changed, but their execution has. And the most important initial advantage—relentless speed—wasn't even in Helmer's original framework.

A note on the source

This analysis draws heavily from YC's perspective, and it's worth acknowledging their position. Many of the companies cited as examples—Legora, Cursor, Happy Robot, Exa.ai, and others—are YC portfolio companies. That doesn't invalidate the framework, but it does mean YC has skin in the game. They're not neutral observers; they're investors backing specific bets about how defensibility works in AI. The principles here are sound, but the examples naturally skew toward validating YC's investment thesis. Keep that context in mind as you read.

Build something worth defending first

YC is direct about this: for early-stage startups, moats are a secondary concern. The first obsession has to be finding a serious customer problem and building something people desperately want. A moat is defensive by nature, and as YC puts it, "you have to have something to defend."

Founders often try to engineer a moat from a business plan. That's backwards. Durable advantages aren't pre-conceived—they're discovered through solving real problems, engaging deeply with customers, iterating on the product, and uncovering unique data and workflows along the way. Worrying about five-year defensive strategy before achieving product-market fit is like trying to defend water in a field before you know where it will pool. Find the treasure first. Then worry about protecting it.

Speed is the real first moat

Before any of Helmer's seven powers matter, a startup's strongest weapon is speed. Varun from Windsor.ai and YC co-founder Paul Graham have both championed this: execution speed is the primary initial advantage for any new venture. It's the one thing a small, focused team has that large incumbents can't replicate.

Look at Cursor, which ran one-day sprint cycles to ship features daily. Compare that to Google or OpenAI, where shipping a new feature can take months or years because of internal bureaucracy, product reviews, and organizational inertia. This differential in execution speed lets a startup gain a foothold, iterate based on real feedback, and achieve product-market fit before an incumbent can respond. Speed helps you discover what's worth protecting. The seven powers are what you build to defend it.

The seven powers, translated for AI

Helmer's "Seven Powers" came from a pre-AI era—Oracle, Toyota, companies like that (https://7powers.com/). But the underlying principles still hold for building lasting value. The concepts of defensibility are timeless; their modern forms have just adapted to AI's unique dynamics. Here's each power translated into an AI context, using YC companies building durable businesses today.

Process power: from assembly lines to reliable agents

Process power comes from building complex operations and methods that competitors struggle to replicate. In the industrial era, think Toyota's assembly line. In the AI era, will it be the painstaking work of creating a reliable, production-ready AI agent?

Anyone can build a weekend hackathon version of an AI tool. That's nowhere close to a system a bank or law firm can actually depend on. Take Casetext for legal work, Greenlight for banking KYC, or Casa for loan origination. They build their moats by pushing past the 80% solution. Their advantage comes from handling edge cases and ensuring consistency. This tedious work—focused, vertical-specific engineering—is exactly what large, horizontal model labs don't want to do. That opens the door for startups to build deep, process-driven advantages.

Cornered resources: from diamond mines to proprietary data

A cornered resource is a uniquely valuable asset that a company can acquire in a way competitors can't. For AI companies, this has evolved beyond physical assets into three categories:

Regulatory and governmental access: Scale AI and Palantir have built moats by navigating the expensive, complex worlds of defense and government contracting. Meeting stringent security requirements, hiring cleared personnel, even building specialized data centers (SCIFs)—these advantages take years to replicate.

Proprietary data and workflows: The modern version of "prospecting" involves forward-deployed engineers who embed with customers to meticulously map their unique, often tedious internal processes. This deep understanding translates into valuable private datasets and, just as importantly, evaluation suites ("evals"). Capturing workflows gives you a cornered data resource and lays the groundwork for deep process power and switching costs that lock in enterprise customers.

Proprietary models: Once seen as the only moat in AI, having a proprietary, fine-tuned model is now just one possible advantage. Character.ai gained an early edge by fine-tuning models to dramatically lower serving costs by up to 10x, creating an operational advantage.

By securing these resources, companies entrench themselves deeply in their customers' operations, making it difficult for those customers to even consider leaving.

Switching costs: from data migration to workflow integration

Switching costs are the pain, effort, and expense a customer faces to move from your product to a competitor's. In AI, this moat has taken on a new, more powerful form.

First is the old SaaS-era moat of data lock-in. Salesforce and Oracle built switching costs by becoming the system of record, where migrating datasets was a painful, multi-year project. This still exists, but LLMs may paradoxically reduce it by automating complex data migration.

The second, more potent AI-native moat is deep workflow integration. Happy Robot and Salient engage in long pilot periods with large enterprises—sometimes six months to a year. During this time, they go beyond importing data to build customized agents woven into the fabric of the customer's specific operational logic. Once this integration is complete, ripping it out to run a bake-off with a new vendor becomes prohibitively expensive. The takeaway: prioritize deep, custom workflow integration over simple data lock-in. The former is growing stronger; the latter may be weakening.

Counterpositioning: from cannibalizing revenue to building better products

Counterpositioning is when a startup adopts a business model or product strategy that incumbents can't copy without damaging their core business. For AI startups, this plays out in two ways:

Business model disruption: Most SaaS incumbents (Zendesk, Intercom) rely on per-seat pricing. This becomes their Achilles' heel in the age of automation. If /when their AI agents succeed, they reduce the number of human employees needed, cannibalizing their own revenue. AI-native startups can adopt usage-based or task-completed pricing instead. Aoka, which builds software for HVAC companies against incumbents like ServiceTitan, captures 4-10% of a customer's operational spend rather than being limited to smaller software budgets (typically ~1%).

Second-mover advantage: AI's rapid pace creates opportunities for second movers who can learn from the first mover's bet and build something demonstrably better. The key is not to copy, but to counter by focusing on a different layer of the stack or a more authentic user benefit.

The legal AI space offers perhaps an emerging example. Harvey, founded in 2022, gained an early lead by focusing heavily on model fine-tuning. They raised substantial capital (over $100M) and built their advantage around creating legal-specific models trained on proprietary data. For a time, this looked like the winning strategy—invest in the model layer, own the intelligence. This tied to an initial focus on leading law firms, created a reputational flywheel.

Then came Legora (a YC company, notably). Rather than compete on model quality, they made a different bet: that the application layer matters more than the model. While Harvey poured resources into fine-tuning, Legora focused on building superior workflow tools, better integration with legal software, and a more intuitive interface for how lawyers actually work. They assumed that foundation models would improve rapidly enough that model-level advantages would erode, but application-level advantages would compound.

The market has yet to validate Legora's approach (of the dozens of legal tech startups that have dropped into this space). Reportedly, law firms that initially went with Harvey are now evaluating alternatives, not because Harvey's AI is bad, but because the product experience doesn't fit their workflows as well. Legora countered Harvey's model-first strategy with an application-first strategy, and in doing so, may have found a more defensible position. This is still playing out, grab your popcorn.

This pattern shows up elsewhere too. Duolingo is a major brand built on gamification, but Speak counter-positions as a tool for people who actually want to learn a language, using LLMs for real speaking practice instead of games. The second mover's advantage comes from watching the first mover's choices and deliberately zigging where they zagged.

Network economies: from social graphs to data flywheels

Network economies occur when a product becomes more valuable as more people use it. Classic example: Facebook. In AI, this principle shows up as a data flywheel.

Every user interaction improves the product. Whether it's a consumer app like Cursor tracking every keystroke to improve code completion, or an enterprise agent capturing usage data to refine workflows, this data feeds back into the system. Evaluations ("evals") drive this flywheel. By systematically analyzing successes and failures, a company can continuously improve its context engineering, prompts, and overall model performance. The result is compounding advantage: more usage leads to better data, which leads to a better product, which attracts more users.

Scale economies: from physical infrastructure to foundation models

Scale economies are cost advantages from large-scale operations. In AI, this moat is most apparent at the model layer. The enormous capital investment required to train a frontier foundation model creates a barrier to entry for labs like OpenAI, Anthropic, and Google DeepMind.

But this advantage isn't exclusive to the labs. The "infrastructure moat" is a repeatable playbook for startups. Exa.ai provides search infrastructure for AI agents. By making a large, early capital investment to crawl substantial portions of the web, Exa built a data asset that can be amortized across many customers. This gives them a cost and infrastructure advantage that new entrants struggle to replicate. Newer YC companies like Channel 3 and Orange Slice are executing this same playbook, showing its viability at startup scale.

Brand: from consumer trust to AI leadership

Brand is when consumers choose a product based on reputation, even when functionally equivalent alternatives exist. AI demonstrates brand power through OpenAI's ChatGPT versus Google's Gemini.

Despite Google having one of the world's most powerful consumer brands and hundreds of millions of existing users, OpenAI moved quickly and decisively. It established ChatGPT as the definitive consumer AI brand, forcing Google into a reactive, catch-up position. In a new category, strong brands can be built on speed and product leadership, capable of overtaking even the most entrenched incumbents.

What to do at each stage

Understanding these seven powers is academic until you put them into practice. The approach differs dramatically for teams still searching for what to build versus companies that have found it and must now defend against competition.

Before product-market fit: ignore moats entirely

If you're an early-stage founder, the directive is simple: ignore the seven powers for now. Focus exclusively on identifying a specific customer with an existential problem—a problem so severe that, as YC describes it, "it's a oh I am not going to get promoted this year maybe I will get fired like this is so painful that I don't want to go to work today."

You're looking for the zero-to-one of problem-solving. Find that acute problem and build the first version of a solution that alleviates it. Worrying about a moat before you have a product people want is premature. At this stage, the only moat that matters is relentless execution speed—the thing that lets you find and solve that problem faster than anyone else.

After product-market fit: build the fortress

Once you've achieved product-market fit and competition inevitably arrives, your challenge shifts. You transition from relying on speed alone to deliberately building durable moats. Analyze your product and market through the lens of these seven powers to identify which forms of defensibility are most natural to your business.

If your product is deeply integrated into enterprise workflows, strengthen switching costs and process power. If your advantage comes from unique user interaction data, accelerate your network economies data flywheel through better evaluation systems. If you're challenging an incumbent, sharpen your counterpositioning by leaning into a different business model or a fundamentally better product experience.

This is where you turn early success into something that lasts, ensuring your hard-won territory can be defended for years.

So now what?

For AI startups, durable moats aren't planned on day one—they're earned through relentless execution and a deep focus on customers. Speed gets you into the game, but long-term value is protected by thoughtfully layering in moats, such as deep workflow integration, proprietary data flywheels, and superior product execution. First, build something worth defending. Then use this framework to protect it with advantages that can last.