Base44 is a vibe coding platform, the kind where you describe an app and it builds it. This week it did something that looks backward at first glance. It launched its own model, Base1, instead of just calling someone else's.
The reason is the most important word in startup strategy right now. Defensibility.
Why Build a Model When Better Ones Exist
Base1 will not beat the frontier on raw capability. Base44 knows that. It trained the model on tens of millions of real user interactions and tuned it for latency, cost, and efficiency on the one job its product actually does. The point was never to win a benchmark. The point was to stop renting the engine of the business from a company that could raise the price or change the terms next quarter.
That is the quiet anxiety running through every AI startup built on top of OpenAI or Claude. If your product is a thin wrapper around a model anyone can call, what exactly do you own? A VC analyst quoted in the TechCrunch piece, Jonathan Userovici, named the three things that make an AI company defensible. Data, distribution, and the tech stack. A wrapper controls none of them. It rents all three.
Base44's bet is to own all three instead. The user interactions are proprietary data. The platform is distribution. And now the model is part of the stack rather than a line item on someone else's invoice. Vertically integrated, in the old language.
The Wrapper Era Is Ending
For two years you could raise money on a clever interface over GPT. That window is closing, and the people closing it are the model labs themselves. Anthropic runs Claude Code. xAI runs its own coding surface. The labs have the model and now they want the application too, with a data advantage no wrapper can match.
When your supplier becomes your competitor, the wrapper is not a business. It is a feature waiting to be absorbed.
I have made this point in different words. If it can be copied, it is already lost. A model call is the most copyable thing in software. The prompt, the interface, the workflow, all of it can be cloned in a weekend by the lab that sells you the model. What cannot be cloned that fast is the proprietary data you have collected, the distribution you have earned, and the tuning that only makes sense for your slice of the market.
That is why Base44 trained on its own user interactions. Nobody else has them. The model is mediocre in general and sharp at the one thing, and the one thing is the moat.
What This Means If You Are Not Building Models
Most companies reading this will never train a model, and they do not need to. The lesson is not about model training. It is about where value sits when the model itself becomes a commodity.
If your AI strategy is a subscription to a tool everyone else also subscribes to, you have bought capability, not advantage. The capability is real and worth having. But it is the same capability your competitor bought, from the same vendor, at the same price. That is not a moat. That is table stakes that feel like progress.
I see this constantly with companies that announce an AI strategy and mean a procurement decision. They bought the same large language model their competitor bought, wired it into a chatbot, and called it transformation. Six months later the competitor ships the identical feature, because they bought the identical thing from the identical vendor. The spend was real. The differentiation was imaginary.
The work that compounds is unglamorous. Collect data nobody else has. Build the relationships that make customers stay through a price increase. Develop the in-house judgment to deploy AI faster and more accurately than a competitor running the same model blind. None of that fits in a press release, which is exactly why it lasts.
The defensible layer is what you wrap around the model that others cannot replicate. Your proprietary data. Your customer relationships and distribution. The judgment of people who know your market well enough to tell good output from confident nonsense. I have argued that distribution is the only moat left, and Base44's move is the same thesis from the engineering side. Own the things that are slow to build, because the fast things are now close to free.
Base1 is a small model from a small company that will not trouble the frontier. As a piece of strategy, it is one of the clearer signals of where this market is heading. The model is becoming the cheap part. The expensive part, the part worth defending, is everything specific to you that a general system can never absorb. Build there, or build on rented ground and hope the landlord stays friendly.