OpenAI and Anthropic's Real Play: Lock Developers In With AI-Generated Code Bases

Samuel Colvin, CEO of Pydantic — the widely used Python data validation framework backed by a $12.5 million Sequoia-led round — argues that OpenAI and Anthropic are executing a deliberate lock-in strategy that has little to do with model quality and everything to do with profit margins ahead of anticipated IPOs. His vantage point is unusually credible: Pydantic sits at the infrastructure layer used by both frontier labs and the developers building on top of them.

The first phase of the strategy is already visible in the aggressive pricing of coding tools. Claude Code and OpenAI's Codex come bundled into $200/month subscriptions even when the actual inference costs run into the thousands of dollars per user. Colvin's read is straightforward: both companies are buying market share at a loss now so they can raise prices once dependency is established. The mechanism for that dependency is the code base itself — if a company uses AI to generate 20,000 lines of code overnight, no human engineer can realistically audit, maintain, or migrate away from it. The AI becomes the only practical tool to manage what the AI built.

Colvin's more provocative prediction is about what comes next: a trajectory or 'intent database.' He expects OpenAI and Anthropic will soon offer corporate subscribers a stored record of every AI-human exchange that produced each line of code — essentially a queryable history of developer intent baked into the code base. Click on a buggy line, and instead of a comment, you'd see the full reasoning chain, model outputs, and human prompts that produced it. That context could dramatically reduce the risk of modifying legacy code, since engineers would know not just what a line does but why it was written that way.

The catch, per Colvin, is the export clause — or lack of one. He anticipates these trajectory databases will be offered for free but made non-exportable, cementing enterprise customers to whichever AI provider they started with. The pitch will be genuinely useful, he concedes, which makes it more effective as a lock-in mechanism, not less. The combination of unmaintainable AI-generated code and proprietary intent metadata could make switching costs for large enterprises almost prohibitive within a few years.