For years, software companies built moats around technology.
A faster database. A better search engine. A proprietary algorithm. The assumption underneath all of it was simple: if the technology was hard enough to build, competitors couldn't easily copy it. The difficulty was the defense.
AI is quietly dissolving that assumption.
Today a startup can build a product on top of frontier models from multiple providers. Tomorrow those same capabilities are available to everyone else, at a lower price, behind a simpler API. When intelligence is rented instead of owned, the advantage you rent is the advantage your competitor rents next quarter.
When intelligence is rented instead of owned, technology advantages become temporary.
Which leaves one question worth actually sitting with: what remains when everyone has access to the same intelligence?
The wrapper debate misses the point
People love dismissing products as "just an AI wrapper."
Sometimes they're right. But the problem was never that a product uses another company's model. The problem is when the model is the entire value proposition.
If a competitor can recreate your product by calling the same API, rebuilding the interface, and copying the prompts, then you don't have a durable advantage. You have a feature. And a feature that fits in a screenshot has a half-life measured in weeks.
This is the same force I described from the engineering side in The Model Myth: when the output is cheap to produce, the output stops being where the value lives. There I was talking about code. The logic doesn't change when you scale it up to a whole product.
Intelligence is becoming infrastructure
Most companies don't differentiate on electricity. Most don't differentiate on cloud hosting. Most don't differentiate on relational databases.
Those technologies matter enormously. A business built on top of them lives or dies by them. They're just not where competitive advantage lives, because everyone has the same access to them. They're the floor, not the ceiling.
AI is moving into that category faster than anyone expected. The model is becoming infrastructure rather than differentiation, and the companies that win will be the ones that build valuable systems around intelligence, not merely on top of it.
The winners won't be the companies built on intelligence. They'll be the ones that built something intelligence happens to plug into.
Where the real moats come from
If the model is the commodity, the moat has to be the thing the model can't supply on its own. A few of those hold up.
Proprietary data. The easiest model to copy is the one everyone can access. The hardest thing to copy is years of accumulated business knowledge and customer data. A competitor can use the same model. They cannot instantly reproduce your dataset, because that dataset is a function of time and operation, not money.
Workflow ownership. The strongest products become part of how work gets done. Replacing them means changing behavior, retraining users, and rebuilding processes. That is dramatically harder than recreating a UI, and the difficulty is the moat. Nobody rips out the thing the whole team has organized itself around.
Operational expertise. A lot of valuable systems aren't valuable because they're intelligent. They're valuable because they reliably solve hard real-world problems. Reliability. Compliance. Security. Observability. Scale. These are the parts that don't fit in a demo and don't compress into a prompt, which is exactly why they're hard to replicate. This is the same point I keep landing on from the reliability side: most reliability problems are organizational problems, and organizations are not something a competitor can fork.
Distribution. A technically superior product that nobody knows about loses to a technically average product that everyone already uses. This was true long before AI, and AI doesn't change it. If anything, a world where everyone can build a comparable product makes the question of who can reach customers the whole game.
Trust. In many industries, customers aren't buying intelligence. They're buying confidence. Confidence that the system is accurate. Confidence that support exists. Confidence that the company will still be here next year to answer for it. Trust compounds slowly and can't be bought with a better benchmark score, which is what makes it durable.
The coming compression
Every major model release follows the same arc.
A capability appears. Everyone gets excited. Competitors copy it. The capability becomes expected. What was a differentiator on Monday is table stakes by the time the next release ships, and nobody remembers it was ever impressive.
That cycle is accelerating. The cost of intelligence is falling. The availability of intelligence is rising. And so the lifespan of any advantage built purely on intelligence is shrinking toward zero. You can win a quarter on a capability. You can't build a company on one.
The better question
A lot of founders are asking the same thing right now:
"How can we build something AI-powered?"
It's the wrong question, in the same way that "how can we write more code" was the wrong question. The useful version is harder:
What advantage remains after everyone has access to the same AI?
Because eventually they will. The model you're building on will be commodity infrastructure, available to your competitor at a lower price and a smaller context-window argument. That's not a risk to hedge against. It's the default future, already mostly arrived.
And when it gets here, the companies that survive won't be the ones with the prettiest wrapper. They'll be the ones that used cheap intelligence to build something expensive to copy: the data, the workflow, the operational track record, the distribution, the trust. The intelligence was always going to be available to everyone. The point was to build the thing it can't hand you.
The one-line version
Intelligence is becoming infrastructure, and you don't win on infrastructure everyone shares. Build the business around something harder to copy than the model, because the model is the part everyone gets.