Models Aren't Products– Even For Developers

OpenAI created Codex, the first model that could generate code from natural language prompts, all the way back in August 2021—before ChatGPT even launched. Now, they’re reportedly in acquisition talks to buy Windsurf, the maker of a popular code assistant that, yes, generates code from natural language prompts. The price? $3 billion. This comes after OpenAI allegedly failed to acquire market leader Anysphere, the team behind  Cursor—another breakout code assistant.

Improving the code generation capabilities of their LLMs, and attracting as many developers to their platform as possible, has been a key focus of OpenAI (and every other AI lab).

Ironically, both Windsurf and Cursor generate code using the underlying foundation models from OpenAI and other labs. So, why is OpenAI trying to acquire them?

Because, as I always say: models aren't products. You need great software products built on top of models to make those models actually useful.

This certainly applies to the vast majority of us who don’t code or don’t know how to optimize context windows with exactly the right chain of prompts. Despite OpenAI being “on the path to AGI,” models alone or even conversational assistants like ChatGPT, don’t offer the right user experience and don’t often solve real problems for every use case or every type of user. Far from it.

But apparently, this is true even for folks that do code. Developers, in fact, have very specific requirements. They demand seamless integration with familiar IDEs ("Integrated Development Environment") and robust support for their preferred coding languages and frameworks. They need tools that maintain deep contextual awareness of entire codebases, enabling precise multi-file edits and proactive suggestions like auto-imports or code optimizations tailored to project conventions. They also require tight workflow integration with tools like GitHub, alongside real-time code completion and specialized AI agents that anticipate needs—such as refactoring codebases or generating API endpoints. A model alone can’t "drop in" and deliver immediate value. Real-world usage comes from thoughtful UX, useful workflow integrations, agentic task completion and true empathy for the user. Windsurf and Cursor have achieved this through flawless product execution, embedding proactive AI that streamlines complex tasks, enhances collaboration, and sets a new standard for developer tools.

If OpenAI acquires Windsurf, they’re getting more than just a polished tool. They’re buying distribution, telemetry, and a direct feedback loop from real-world users. It’s a strategic bet on a vertical: going beyond the model and building the “last mile” to the user. And in doing so, they will own the user experience, the data, and the actual workflows.

In my experience, this is exactly where moats in software have always come from.

We have been told that foundation models are the future. That a single, super-powerful model could disrupt entire industries with scale, generalization, and elegance. As an investor in OpenAI, I appreciate the very necessary platform layer that these foundation model companies are building.

But, I believe that the next quantum leap in AI will come from building insanely useful and usable products that impact the 8 billion people who don’t currently truly benefit from AI today. 

At Axiom, we call this: AI for the Real World.

If you’re a founder building in AI and thinking at this scale, let’s connect.