Pydantic’s Samuel Colvin warns OpenAI and Anthropic will lock you with coding-intent databases

Pydantic CEO Samuel Colvin thinks OpenAI and Anthropic are quietly preparing a new kind of lock-in: not just access to a model, but access to a database of “coding intent.” In Colvin’s view, once enterprises build huge codebases with AI help, the vendor that knows the history of how code was produced can become the vendor you cannot easily leave. That is the pivot he sees coming, and it matters because it reframes what success means for these labs as well as what risk looks like for the companies signing long-term subscriptions.

Colvin’s argument is blunt. He expects OpenAI and Anthropic to store “the traces or trajectories of the full exchange between users and the model as it’s writing code,” turning that into a lookup database where customers can find the intent behind specific lines. He also expects the catch: they would likely “give you that for free, but you can’t export it.” Translation: the value grows inside the vendor’s system, while portability shrinks.

To understand why Colvin believes this is the next play, you have to follow the economics of frontier model businesses, which he says have moved from revenue-first to profit-margin-first. He tells Business Insider that a year ago, what mattered most was revenue, including anything that could drive usage of inference. But as the labs work toward IPO goals, margin becomes more important. Colvin’s framing is that competing purely on model quality drives huge training spend, plus cheap inference becomes mandatory if you want margin. So, instead, labs search for ways to lock in customers that do not depend on constantly out-training everyone else.

That is where today’s AI coding offerings fit into the story. Colvin points to “Claude Code and Codex,” including the “discounts - $200 a month subscriptions when you are actually spending maybe thousands on inference with that subscription.” His point is not that the discounts are inherently predatory. It is that they are a growth lever. Grow market share. Increase usage. Then, once usage becomes embedded, a second stage becomes possible: with huge codebases, corporate customers have to keep using the same services from Anthropic and OpenAI, and pricing power can follow.

Colvin then adds a practical constraint that enterprises feel immediately once they start generating large volumes of code with AI agents. If AI generates “20,000 lines of code overnight,” a human maintainer may not realistically be able to understand, audit, and keep that codebase alive without the same tooling that produced it. He describes a reality check for human-only maintenance: yes, you can use models to help fix issues, but maintaining and understanding what the code is supposed to do at the detail level becomes harder for humans. In that setting, the company that provides context about the code becomes a recurring dependency, not a one-time generator.

The proposed database of coding intent is supposed to solve a problem that shows up as codebases get larger and comments get messier. Colvin uses an example: imagine a software bug tied to “some odd behavior on some line of code.” If the code author, whether human or AI, left an “helpful explaining comment,” the team might be able to reason about intent. But too many comments can become unreadable. His alternative is to click on a line and see the full exchange that produced it: the inputs from the human, the model’s reasoning, and everything that shaped why that line exists. In his view, this gives a “much richer understanding of the intent” behind a codebase, and it makes changing that line “lower risk,” because you can tell whether something is a bug or an intended behavior.

Colvin’s thesis also highlights a second-order dynamic that boards and procurement teams should notice: the “database of trajectories” and “locking you in” are not just about convenience. They are about switching costs that scale with enterprise size. Once an organization accumulates the full history of exchanges tied to particular code lines across many sprints, the stored context becomes operational knowledge. If it is non-exportable, moving the coding layer becomes more than migrating an app. It becomes losing access to the institutional memory of how your AI-generated code was built.

And while Colvin presents this as his “personal guess,” it lines up with the incentive structure frontier labs face as they balance IPO readiness, margin targets, and customer growth. For companies choosing vendors for AI development tools, the question shifts from “Which model writes the best code today?” to “Which vendor holds the keys to the audit trail of how that code was produced tomorrow?” If Colvin is right, this is where the next lock-in wave could land: in the background systems that capture intent, not just the front-end chat experience where code gets generated.