Why Cursor Built Its Own Model (It's Not About Coding)

By Sequoia Capital

Categories: VC, Startup

Summary

Cursor built its own AI model not for general coding, but to hyper-specialize on a single task: software engineering workflows within their product. By allocating all model weights to one use case, Composer 2 costs an order of magnitude less than Opus while delivering superior performance—a strategy that redefines competitive moats beyond just being an application layer.

Key Takeaways

  1. Think of model weights as finite storage bits. Rather than building general-purpose models, allocate 100% of model capacity to a single, high-value task. This specialization reduces model size and inference costs by an order of magnitude while improving task-specific performance.
  2. Reframe the competitive advantage: don't optimize for coding or programming broadly—optimize for one specific workflow (e.g., software engineering inside your product only). This constraint becomes your moat and enables you to serve smaller, cheaper models than competitors.
  3. Cost structure is a defensible advantage. Composer 2 is order of magnitude less expensive than Opus because specialization allows serving smaller models. Lower unit economics enable aggressive pricing and customer acquisition that generalist competitors cannot match.
  4. Building your own foundation model isn't about becoming a model company—it's about owning the constraint that gives you unfair advantages. For application companies, custom models solve the problem of general-purpose models wasting capacity on irrelevant tasks.
  5. Question whether your startup needs to compete on breadth or depth. If you have a narrow, well-defined use case with clear leverage (like in-IDE coding assistance), custom model training beats licensing generic models on cost, latency, and performance simultaneously.

Related topics

Transcript Excerpt

What was the impetus for Cursor to lean so heavily into Composer 2, and how existential is it for you to become not just an application company, but also a foundation model company yourselves? The reason why we started looking into training our own models is you can sort of think about the model as sort of like a storage drive. It has certain amount of bits that it can store in its weights. And the idea is very simple. We care about only one task. We don't even care about coding or programming necessarily. We care about software engineering inside Cursor and inside Cursor only. And so, what if we were to allocate all of the bits of information that can be stored inside a model weights to that one particular task. Also, as people may have noticed, Composer is order of magnitude less expensi…

More from Sequoia Capital