Five ways to build proprietary data.

Last week I argued a data moat only matters if it compounds. But before it can compound, you have to build the dataset. That is the harder question, and here is what actually worked, from companies we partnered with from their first days.

1) Earn it intimately, one customer at a time. Navina, an AI copilot for primary-care doctors, helped them treat patients better. Practice by practice, that trust earned it access to years of medical records, and every new customer made its models sharper.

2) Go where no one else is willing to go. Alice (Formerly ActiveFence), an AI safety company, spent years in the darkest corners of the internet building Rabbit Hole, what it calls the world's largest store of adversarial intelligence. Knowing how real harm actually operates is what lets it protect billions of users and the largest AI models.

3) Capture the physical world no one bothered to digitize. Limitless Labs automating CNC machining, went to the factory floor and gathered and labeled the real machining data that lived nowhere else, then trained physics-based AI on it.

4) Harvest the public signal at a scale no one else reaches. Onfire AI, a go-to-market platform, built the engine to read the open web where customers reveal what they want, and turn millions of those signals into who is buying what, and when.

5) Create data that never existed. Protai, an AI drug-discovery company, used proteomics to map disease at the protein level and built the largest, most diverse proteomic database in the world, data that never existed before.

None of these stopped at owning the data. Each spun its dataset into a data flywheel, cheaper to feed and more valuable with every turn. And the flywheel, not the data, is what becomes the moat. Grove Ventures