Apple trained a China-specific model with Alibaba's help
Apple quietly built a separate model for the Chinese market in partnership with Alibaba — the clearest example yet of a global product being split into jurisdictional variants because one regulatory environment will not accept the other's system.
Companies have localised products for decades — language, payments, content rules. Training a separate model is a different order of divergence: it is not the interface that differs but the system underneath, with its own training data, its own behaviour and its own approval path.
The requirements make it close to unavoidable. Chinese regulation requires generative models to be registered and approved, with content constraints that are not adjustable settings on a model trained elsewhere. Building locally with a domestic partner is the compliant route, and Alibaba brings both the capability and the regulatory relationship.
The cost is a permanent fork. Two models means two evaluation regimes, two safety pipelines, two update cadences and two sets of behaviour to reason about — and the divergence compounds with every release rather than converging.
It is the same structural pressure visible from the other direction this week, where Washington is tightening who may receive advanced chips through which corporate structures. Both jurisdictions are asserting that AI systems touching their markets must be built under their rules, and the resolution is not convergence. It is two stacks.