Open source won most of modern software infrastructure because it aligned with how software value was created. Linux and PostgreSQL spread. Kubernetes and the web stack spread. Developers could adopt them, extend them, and build around shared standards. Distribution was cheap, and reuse created compounding network effects.

AI changes that equation.

Model weights are important, but they are not the full product. Real value often sits in orchestration and evaluation, then in deployment, safety controls, workflow integration, and feedback loops. Inference also carries an ongoing marginal cost. Instead of shipping code once, you pay to run intelligence repeatedly.

That is the break from classic software. A database can be downloaded and hosted, then improved by a company with a conventional engineering team. A frontier model depends on training runs, GPU supply, data pipelines, eval infrastructure, serving economics, and constant behavior management. The artifact matters, but the operating system around the artifact matters too.

That shifts the strategic question. In classic open source software, "open code" was often enough to create a durable ecosystem. In AI, openness can mean different things:

  • open weights
  • open training pipeline
  • open evaluation methods
  • open deployment stack
  • open interfaces and portability

Most offerings are selectively open. That is not automatically bad, but it means founders and buyers need to be precise.

The practical issue is not ideology. It is control:

  • Who controls model behavior over time?
  • Who owns the user and workflow?
  • Who captures improvement data?
  • Who can switch vendors without a rewrite?

Open versus closed in AI is not one decision. It is a stack of decisions across product and model choices, plus infrastructure and distribution.

This series will avoid slogans and focus on tradeoffs. Closed systems can produce better product experiences and accountability. Open systems can create trust and adoption, plus portability and ecosystem pressure. Both can be strategically correct, depending on the layer and market.

The useful question is not whether a company is "open" or "closed." It is what part of the system must be inspectable, what part must be dependable, what part must be portable, and what part should be tightly integrated enough that customers simply get the outcome.

The right answer starts here: stop asking who is morally right, and start asking where durable advantage is actually built.

This is part 1 of 10 in Open vs Closed AI.