Apple builds China-only AI with Alibaba as US weighs new AI rules

Apple just made a decision that tells you more about the geopolitics of AI than any policy paper could. When Apple Intelligence launches in China, it won't run on Google Gemini, and it won't run on a homegrown Chinese model licensed off the shelf. It'll run on a large language model Apple built itself, reportedly with engineering help from Alibaba. Everywhere else in the world, Apple's strategy has been to partner rather than build — Gemini in most markets, OpenAI for certain features. China is the one place where Apple decided the stakes were too high to outsource entirely.

Why does this matter beyond the headline? Because it's a tacit admission that no existing arrangement worked. A foreign model would run into regulatory approval problems and likely nationalist backlash. A fully Chinese third-party model would mean handing a competitor deep access to the data and design of Apple's most personal product layer. So Apple split the difference: its own model, trained with local partnership, to satisfy Beijing's data and content rules while keeping the crown jewels — the integration layer, the user experience, the actual product — under its own control. Reuters' reporting suggests this wasn't a quick pivot either; it's been in motion for a while, which tells you Apple saw this coming long before Apple Intelligence launched anywhere else. That's a template other Western tech companies operating in China will study closely, especially anyone whose AI strategy currently assumes a single global model can serve every market and every regulator.

Meanwhile in Washington, the conversation is moving in the opposite direction — toward less certainty, not more. Congressman Don Beyer's comments about Congress needing to rewrite its rulebook aren't new in substance, but the timing is telling. AI capability is compounding faster than any committee hearing schedule can track, and the result is a policy vacuum that individual states are rushing to fill on their own terms. That's produced a genuinely dangerous mismatch: federal deregulation instincts colliding with a patchwork of stricter state rules, at precisely the moment when AI-powered cyberattacks are starting to move at machine speed rather than human speed. I find this concerning not because deregulation is inherently wrong, but because a fragmented response to a borderless threat is close to no response at all. Attackers don't care which state's rules apply.

Put those two stories together and you get a rough sketch of where AI governance actually sits right now — not in some future regulatory regime, but in ad hoc, company-by-company and state-by-state improvisation. Apple is building its own compliance infrastructure because no government gave it a clean answer. American lawmakers are debating whether to modernize rules while attackers already exploit the gaps. The interesting question isn't whether governments will eventually catch up — they will, eventually. It's how much gets built, and how much damage gets done, in the meantime.