Your data moat is a mirage.
I see it in pitch decks every single day. Founders, bright-eyed and armed with impressive charts, tell me about their "insurmountable data moat." They believe the more data they collect, the better their AI model becomes, and the more unbeatable their company will be. It’s the classic data network effect, a concept preached like gospel in Silicon Valley. And for many industries, it works.
But not in healthcare. Not in the brutal, regulated, high-stakes world of AI diagnostics.
I’ve been in the trenches of Silicon Valley for over a decade. I’ve built and sold two companies—RemoteTeam to Gusto and MovieLaLa to Gfycat. I’ve written over 200 angel checks into companies that are now household names, like Anthropic, OpenAI, Scale AI, and Hugging Face. I’ve seen what it takes to win, and I’ve seen what leads to spectacular failure. And I can tell you this: the conventional wisdom about data moats is not just wrong in healthcare AI; it’s a dangerous delusion.
Most founders think building a great AI model is enough. They're wrong. The uncomfortable truth is that what it really takes to succeed in healthcare AI has almost nothing to do with the size of your dataset. It’s about how you get it.
The Great Illusion: Why Your Data Moat is Useless
The myth of the data moat is seductive. It goes like this: more data leads to a better model, which creates a better product, which attracts more users, which generates even more data. It’s a beautiful, self-perpetuating cycle that looks fantastic on a slide. I get it. It’s what we’re all taught to chase.
But this model shatters the moment it collides with the reality of clinical medicine.
The Clean Data Fallacy
Here’s the first hard truth: your model, trained on pristine, perfectly labeled public datasets, is a pampered house cat. Real-world clinical data is a feral beast. It’s a chaotic, snarled mess of inconsistent formats, missing fields, scanned PDFs, and hidden biases from a dozen different EMR systems that barely talk to each other. Your elegant algorithm will choke on it.
I remember meeting a team a few years back. Two PhDs from a top university, incredibly smart. They had built a phenomenal algorithm for detecting a specific type of cancer from imaging. Their accuracy on the public benchmark dataset was north of 99%. They thought they were unstoppable. I passed on the investment. Why? Because their entire model was built on a fantasy. They had never tried to run it on the messy, incomplete, and utterly unpredictable data that comes out of a real hospital’s workflow. Their model wasn t just useless; it was a liability.
The FDA Brick Wall
Let’s say you solve the messy data problem. Now you face a bigger one: the FDA. The regulators in Washington don’t give a damn about your model’s performance on a generic dataset. They have one question, and one question only: does your product work safely and effectively for its intended use in a specific, representative patient population under real-world conditions?
This means you can’t just hoard data. You need to conduct rigorous, expensive, and painfully slow clinical trials. You need to prove that your AI doesn’t just work in a lab, but that it improves patient outcomes in a busy, chaotic emergency room or radiology suite. That requires a targeted, strategic data collection effort from day one, designed to meet the FDA’s exacting standards. Your generic "moat" is worthless here.
The Integration Nightmare
Even if you get FDA clearance—a massive achievement in itself—you’re still at the starting line. Now you have to sell your product. This means integrating with hospital IT systems, a process so notoriously difficult it has become a running joke in the industry. Your brilliant AI is just another piece of software that has to fit into a dozen legacy systems, each with its own quirks and gatekeepers.
Your model’s accuracy is a feature, but the sale is won or lost on workflow. If your tool makes a clinician’s life even 1% harder, they will not use it. And if they don’t use it, you don’t get the data, and your flywheel never starts spinning. The product is dead on arrival.
The Real Moat: Building a Clinical Data Engine
So if the traditional data moat is a trap, what’s the answer? You have to stop thinking about a static "moat" and start thinking about a dynamic "engine." The goal isn’t to collect a pile of data. The goal is to build a system—a clinical data engine—that continuously generates proprietary, defensible, and regulatory-grade clinical data through the very use of your product.
This is the real moat. It’s not a wall you build; it’s a factory that produces your most valuable asset. Here’s how you build it.
Step 1: Go Painfully, Uncomfortably Niche
First, you have to forget about boiling the ocean. Don’t build a tool that promises to "read all X-rays." That’s a recipe for disaster. Instead, find one specific, agonizingly painful problem for one specific type of clinician and solve it better than anyone else.
I invested in a company that did this perfectly. They didn’t try to build a general radiologist AI. They built a tool that did one thing: help pediatric radiologists detect a rare but often-missed type of wrist fracture in children. It was a tiny niche. But it was a high-stakes problem where a mistake could lead to lifelong disability. They focused all their energy on becoming the undisputed best in the world at that one task. They dominated that beachhead, and only then did they begin to expand. That’s how you win.
Step 2: Design for Data Capture, Not Just Diagnosis
This is the most critical mindset shift. Your product’s primary job is not just to provide a diagnosis. Its primary job is to create a feedback loop. The user interface has to be masterfully designed to seamlessly capture the clinician’s agreement, their corrections, their annotations, and their ultimate diagnosis.
Every single time a doctor uses your tool, that interaction must become a valuable, structured data point that feeds back into your system. This is how you generate a proprietary dataset that no competitor can ever replicate. They can’t buy it, and they can’t scrape it. It is born directly from the value your product creates. This is your engine.
Step 3: The Human-in-the-Loop Flywheel
When you put it all together, you create a powerful flywheel. Your initial model, even if it’s not perfect, is good enough to get you in the door because it solves a painful, niche problem. A clinician uses it. The product’s brilliant design captures their feedback—the essential "human in the loop." That new, proprietary data is used to retrain and improve the model. The now-better model provides even more value to the clinician, which leads to more usage, which generates more data.
This is a compounding advantage. It’s a self-improving system where your product gets smarter with every single use. This flywheel is the real, defensible moat in healthcare AI. It’s an advantage that a competitor starting with a static, generic dataset can never, ever catch up to.
Stop Chasing Moats, Start Building Engines
The old Silicon Valley playbook of hoarding data is a trap. It will lead you to waste millions of dollars and years of your life building something that will never survive contact with the real world of healthcare.
The future of AI in diagnostics belongs to the founders who understand this. It belongs to the builders who ignore the siren song of the generic data moat and instead focus on building a clinical data engine.
It’s a harder path. It’s slower. It requires a deep, almost obsessive focus on a narrow problem and a relentless commitment to workflow and user experience. But it’s the only way to build a company that lasts. It’s how you win in the brutal, regulated, but incredibly important world of healthcare AI. Stop admiring the problem and go build the engine.
Frequently Asked Questions
What experience informs this perspective?
This perspective comes from over a decade of building companies in Silicon Valley, two successful exits (RemoteTeam to Gusto, MovieLaLa to Gfycat), and investing in 200+ startups including Anthropic, OpenAI, and Scale AI. I write about what I've lived.
Do all experts agree with this view?
No, and that's fine. The best ideas in business are often contrarian. I share my perspective based on my experience and data, but I encourage you to seek out opposing viewpoints and form your own conclusions.
How can I apply this thinking to my own situation?
Start by identifying the core principle behind the opinion, not the specific example. Then ask yourself: does this principle apply to my context? If yes, test it in a small, low-risk way before going all in.
What's the most common pushback you get on this?
People often push back by citing exceptions or edge cases. And they're usually right that exceptions exist. But building a strategy around exceptions rather than patterns is a losing game for most founders.