As the models become more and more commoditized, the differentiation moves further up the stack.

First, the infrastructure and harness that sits on top — orchestration, memory, identity, context, task management and execution — all of that intricate tooling can dramatically impact how the models perform and what they are capable of.

Products like OpenClaw or Cursor have proven this in different ways, and there is immense value in having agent harnesses that draw out extraordinary things from the models.

Claude Code and Codex are the most obvious examples, but people get confused because those are built by the companies that also produce the models. But the product teams at Anthropic and OpenAI are working with the same models that third-parties have access to, the same way that product teams at Google are building with the same publicly-available Gemini models.

And so this all begs the question: do the labs that develop the models necessarily need to be the ones that create the infrastructure and the applications? Do they have any inherent technical advantage at these levels of abstraction that sit on top of the models?

They have capital and people resources. They have brand power. They get early access to new models and can build ahead of public releases. They get a head start, but does anything technically separate their ability to create great products from anyone else building on the same models?

It seems that there is room in the market for new players to emerge who compete solely at the application and infrastructure layers and do not produce their own models. They can accrue value from ingenuity in UX, deeply understanding and serving particular types of customers, and building smarter harnesses.

And since they aren't developing the models, these are much more lightweight businesses that more closely resemble "normal" software companies: relatively low capex, high gross margins.

There is no structural reason why you can't have a multi-billion dollar agent business that does not produce its own models.

It's everything but the model that might end up having the most value as the market matures.