The Data Moat Didn't Move. It Changed Shape.

The founders who are scared of AI are usually scared of the wrong thing.

They see foundation models improving quarterly and assume the gap closes eventually — that if GPT-whatever can do what your product does, you're done. The logic sounds tight. But it's pointed at the wrong layer.

I ran into this directly while scoping the AI layer for a legacy vertical SaaS platform I'm rebuilding from the ground up — decades of accumulated customer workflows, being rearchitected around AI rather than patched. The question I kept asking myself wasn't "what can the model do?" It was "what does the model not know?"

The answer told me everything about where the moat actually sits.

The layer that gets commoditized

Models converge toward parity. This is normal, and it's happened in every prior infrastructure cycle: databases, cloud compute, mobile operating systems. The commodity layer always drops in cost until it's infrastructure.

That's what's happening with foundation models. The gap between GPT-4 and its competitors in 2023 was enormous. The gap between frontier models today is much narrower — and it keeps narrowing. By the time you ship a product differentiated purely by model quality, the model quality advantage is usually gone.

Building your competitive advantage on the model layer is building on the part that commoditizes fastest.

The layer that compounds

Context is different.

Context is decades of customer workflows encoded in a system that's been in production long enough to have seen almost everything. It's knowing how a specific regional variant of a process gets handled when three conflicting jurisdictional rules collide. It's the edge cases that took a decade to surface, the local variations that aren't documented anywhere publicly, the workflow patterns that experienced users just carry in their heads.

No foundation model trained on the public internet has this. It's not in any training dataset. A competitor who enters the market tomorrow — even one with access to the same AI infrastructure — doesn't have it either.

That's the context layer. And it compounds in ways the model layer doesn't.

Three properties that make it durable:

Feedback loops from daily operations. Not surveys, not NPS scores — the actual in-product decisions customers make every day. Each one is a labeled data point about what works in this domain. At scale, across thousands of active users, those signals accumulate fast.

Domain edge cases that don't exist in public data. Every vertical has these: scenarios too specific to appear in general training data, too important to ignore in production. Any mature vertical has hundreds of them. Building a system that handles them correctly is expensive. It's also a moat.

Historical depth that can't be reconstructed. You can't buy decades of domain-specific workflow patterns. You can't scrape them. You can only accumulate them by operating in the market that long. That asymmetry doesn't go away because AI gets better — it gets more valuable because now you can actually use it.

The right question

The question "will AI commoditize us?" is almost always the wrong question. It focuses on the model layer and treats context as an afterthought.

The right question is: what does the AI know about your domain that no one else's AI knows?

If you're a vertical SaaS company with years of operational data, real customer feedback loops, and domain-specific edge cases encoded in your system — you're not under threat from AI. You have an asset that becomes more valuable as AI infrastructure gets cheaper, because cheap inference is exactly what you need to put your context to work.

The moat didn't move. It changed shape. From features that competitors couldn't build fast enough, to context that competitors can't accumulate at all.

The builders who understand this aren't scared of better models. They're building the dataset that makes the model useful — and that's the part nobody else can copy.