Models move fast. Moats compound.
Why the best AI companies aren’t chasing benchmarks—they’re building unfair advantages across data, distribution, and workflow
The model race is loud.
Bigger, faster, cheaper — every week, a new benchmark.
But here’s the truth: models are becoming commodities. Moats are not.
And if you're betting on models alone, you're already behind.
The most enduring value in AI won’t come from building a better LLM.
It’ll come from building a compounding advantage — one that improves with every user, every interaction, every loop.
Step 1: Own the weirdest data
As I wrote in last week’s post, the next $10B AI company probably won’t win on model architecture.
It’ll win on access — to messy, overlooked, irreplaceable data.
The kind that’s:
Hard to collect
Expensive to label
Boring to most people
Invaluable once structured
📊 Examples:
Forklift telemetry logs
Permit workflows across 1,000 localities
Speech-to-text transcripts from care coordinators
Footage from wildfire towers in rural zones
These aren't datasets you can scrape or simulate.
You have to embed, serve, and earn them.
Step 2: Capture behavior at the edge
In [Post 2], we talked about edge AI — not just as a deployment strategy, but as a data strategy.
The edge is where behavior happens before it's filtered.
It's where urgency, raw signal, and defensibility live.
Whether it's:
A nurse entering post-op notes
A field technician recording a mechanical fault
A wildfire detection system spotting smoke in real time
The startups that control these interfaces aren't just gathering data —
they’re building moats others can’t cross.
Own the edge, own the stream.
Own the stream, own the compounding loop.
Step 3: Embed the loop between data and product
Moats form in the loop — not the launch.
The best AI companies don't just process data.
They feed on it — tightening feedback loops so every user interaction makes the product smarter.
📚 Examples:
Every route driven by a delivery van improves its routing model
Every document edited in a writing tool trains a better summarizer
Every diagnosis assisted by a clinical copilot refines treatment accuracy
This is how models get stickier even as the underlying tech commoditizes.
🧱 A New Mental Model: The Moat Stack
When I evaluate AI companies, I don’t ask “How big is the model?”
I ask “How deep is the moat?”
Here's the lens I use:
Most founders show me their model.
The best ones show me how it compounds.
🧠 For founders: This is your unfair advantage
You don’t need the biggest model.
You need the tightest loop.
Build a system that:
Captures signals others can’t access
Improves itself with every interaction
Embeds deeply into workflows people already trust
That’s how you create a moat — not just a product.
🎯 For investors: A sharper lens on defensibility
Benchmarks will come and go.
Moats? They compound.
The sharpest investors I know are asking different questions:
Where is behavior first captured?
Is AI embedded at the moment of decision?
Is the dataset exclusive and hard to replicate?
Does usage make the model smarter over time?
Is AI the core value — not just a feature?
Because if OpenAI drops a better model tomorrow...
the right companies don’t just survive — they get even stronger.
💬 I'd love to hear:
What’s the smartest weird data moat you’ve seen lately?
Please reply and share if you can — I’m collecting examples.