When generative AI first broke into the mainstream, companies like OpenAI and Anthropic were understood primarily as infrastructure providers. Developers were encouraged to build on top of them, with the promise that AI models would be the foundation layer of a vast new ecosystem of applications. But today, these same companies are climbing further up the stack.
Take OpenAI’s recent release of Sora2, a consumer-facing app for video generation. What was once just a raw capability (text-to-video) is now packaged into an end-user experience, competing head-on with startups that thought they’d have room to build applications. Similarly, Anthropic has launched Claude Teams, not just offering API access to Claude models but delivering a ready-made productivity suite for enterprises.
Think of the model companies as farms. They used to sell ingredients to restaurants — the startups — who turned them into meals. Now the farms are running restaurants too. To stand out, you can either cook better with the same ingredients or source ones no one else can.
This raises a key strategic question: how can startups build defensible businesses when their infrastructure providers are also their fiercest competitors?
Our answer: by planting seeds in walled gardens of data. In this case, walled gardens of data are domains where access to information is restricted, proprietary, and highly valuable — and where exclusivity itself creates a moat. These datasets are typically:
With that definition in mind, let’s review two examples: VLex in law and OpenEvidence in medicine.
VLex, a legal software company in Spain, began in 2000 as an ambitious effort to “revolutionize legal information access” by building a comprehensive legal content platform and applying new technologies to legal research. Spanish court decisions, statutes, and regulations were historically scattered across fragmented regional jurisdictions and often not available in machine-readable formats. Over the years, VLex systematically acquired, licensed, and digitized this content — effectively creating one of the most comprehensive legal databases in Europe. The result is something akin to LexisNexis + Westlaw + Bloomberg Law, specifically for Spanish legal history.
By the time generative AI models became viable, VLex had already amassed a proprietary corpus of legal data spanning decades of judgments and commentary. This gave it a defensible wedge to build AI-native legal research tools: unlike general-purpose models, VLex’s system can actually reason over authoritative, complete, and current legal texts. Its moat isn’t the model — it’s the painstakingly assembled dataset that no one else has.
Said differently: a lawyer crafting the best possible brief or argument needs access to every legitimate matter of precedent. A general-purpose model — even one as powerful as OpenAI’s — might produce a sophisticated-sounding legal argument, but missing even a few critical historical cases could be the difference between winning and losing.
If the stakes are high in law, they’re even higher in medicine. OpenEvidence pursued a parallel strategy to VLex, but in healthcare. While health information is plentiful online, most of it is unvetted or consumer-grade (think WebMD articles or forum posts). Clinicians, by contrast, rely on peer-reviewed literature, systematic reviews, and clinical guidelines — much of which sits behind paywalls like Elsevier or is restricted to medical institutions.
OpenEvidence spent years building partnerships, licensing agreements, and ingestion pipelines to create a structured database of vetted medical research. With that foundation, its AI can answer complex clinical questions with evidence-backed precision, rather than hallucinating or relying on incomplete public data. In medicine, where trust and accuracy are existential, this walled garden is not only a moat but also provides a far superior user experience to that offered by general-purpose models. If you’re researching a personal ailment, you’d much rather get scientifically grounded advice than spiral into WebMD doom-scrolling.
These stories show the power of owning unique, hard-to-access data. But the opportunity doesn’t stop at law or medicine. Across industries, fragmented datasets remain unclaimed — waiting to be cultivated into walled gardens that could anchor the next wave of AI-native companies. Let’s examine a few.
The model companies will always command bigger models, more compute, and more distribution–those are hard games for startups to win. But there’s an opening in ecosystems where high-quality data has historically been fragmented, sensitive, or difficult to access — places where sovereignty and trust matter more than raw capability.
Building new walled gardens isn’t easy: it requires significant upfront investment and often painstaking groundwork, striking deals across companies, governments, and institutions. Yet when it works, the result is nearly impossible to replicate, offering one of the rare durable edges in an increasingly competitive AI landscape.
Are you building a new walled garden? We’d love to hear from you.