Open weight models won the boring half of enterprise AI, and the boring half is most of it.

Ask a European enterprise architect which model they use and you now get two answers. A hosted frontier model for the small number of tasks that genuinely need it, and a self-hosted open weight model for classification, extraction, summarisation, routing and the long tail of internal tooling.

That split did not happen because open weights caught up on the hardest benchmarks. It happened because the tasks that dominate real deployments never needed the hardest benchmarks.

Three forces pushed it

Cost predictability came first. A per-token bill that grows with adoption is a difficult thing to defend once a tool becomes popular internally. A fixed cluster is easier to budget even when it is more expensive in aggregate.

Data residency came second, particularly for anything touching employee or patient records. Third, and least discussed, is version stability. A hosted model that changes underneath a validated workflow creates revalidation work that nobody scheduled.

What the frontier still owns

Long-horizon agentic work, hard code generation, and anything requiring genuine multi-step planning remain clearly better on the largest hosted models. Teams that pretend otherwise end up building elaborate scaffolding to compensate for a weaker model, and the scaffolding costs more than the tokens would have.

The honest architecture

Route by task difficulty, measure the routing, and revisit it quarterly. The organisations doing this well treat model choice as a procurement decision with a review cycle, not as an identity.