The AI industry is obsessed with one question: Who has the best model?
But a recent SemiVision Research essay, “AI’s Next War Is Not About Better Models—It Is About Who Controls Them”, makes a compelling case that another question may matter much more in the long run:
Who controls the infrastructure on which AI products are built?
Today, most AI products effectively rent intelligence. They call an API from OpenAI, Anthropic, Google or another provider and get increasingly capable models without having to operate the underlying infrastructure.
For most products, that is a great trade.
But as AI moves from individual features into the core of products, these APIs also become architectural dependencies. Prompts, evaluations, agent workflows and user experiences increasingly evolve around the behavior of a particular model.
Switching providers may eventually be much harder than changing an API endpoint.
Open models create optionality
This is where open-weight models become strategically interesting.
The comparison with technologies such as Linux, PostgreSQL and Kubernetes is useful. Their importance was never simply that they were free. They gave companies an alternative to depending entirely on proprietary platforms.
Open AI models could play a similar role.
Companies may be able to run them across different clouds, on dedicated infrastructure or in environments where data cannot leave the organization. More importantly, credible alternatives create leverage even for companies that never operate a model themselves.
But open does not mean independent. Running models still creates dependencies on GPUs, inference software and cloud infrastructure. The dependency simply moves to another layer.
A product strategy question
That makes model choice increasingly relevant to product strategy.
Instead of asking only Which model performs best today?, product teams should also ask:
How difficult would it be to replace this model in two years?
The answer does not have to be “easy.” Deeply integrating one provider might create a substantially better product, and that can be worth the lock-in.
The important part is making that trade-off consciously.
Because the best model will keep changing.
The more durable competitive advantage may therefore not come from choosing today's winner, but from deciding which parts of your intelligence stack you are comfortable renting — and which ones might eventually become important enough to control.
Inspired by SemiVision Research's “AI’s Next War Is Not About Better Models—It Is About Who Controls Them”, published July 25, 2026.