The U.S. wants to contain China’s AI. Silicon Valley keeps using it

There's a delicious irony buried in AI's culture-war rhetoric this week. Anthropic's Tarun Chhabra, once an export-control architect under Biden, singled out China's Zhipu for allegedly distilling capabilities from American frontier models — a warning that slots neatly into Washington's preferred story about a one-way flow of innovation from Silicon Valley to Beijing. Hours later, Mira Murati's Thinking Machines, flush with $2 billion in funding and her OpenAI pedigree, revealed that its debut model, Inkling, leans on DeepSeek-V3's architecture and uses synthetic training data from Moonshot AI's Kimi K2.5. So much for the one-way street.

Distillation itself isn't some shadowy new weapon. It's a decade-old machine learning technique — a big teacher model generates outputs, a smaller student model learns to mimic them cheaply. Every major lab does some version of it now, from OpenAI and Google DeepMind to Alibaba and Tencent. What's changed is the politics: distillation has been rebranded, often carelessly, as something close to IP theft whenever a Chinese company is on the receiving end, even as American startups quietly do the exact same thing with Chinese open-weight models like Qwen, Hunyuan, and DeepSeek.

And the entanglement runs deeper than one startup's training pipeline. Apple just got Cyberspace Administration approval to ship Apple Intelligence in China running on Alibaba's Qwen and Baidu's Ernie — two companies the Pentagon has formally tagged as Chinese military-affiliated. That means iPhones carrying Chinese-government-approved models will be crossing borders in the pockets of ordinary travelers, while Washington simultaneously tightens API access and export rules around its own frontier labs.

The practical result is a policy contradiction nobody in DC seems eager to name. Export controls still work well on chips and fabrication equipment — you can physically block a shipment. They work far worse on research papers, open-weight checkpoints, and synthetic datasets that flow freely across the internet. Models like Zhipu's newly released GLM-5.2 are out there, openly licensed, with no meaningful gatekeeping, while American firms build ever-higher walls around Claude and GPT.

What's emerging isn't a single frontier that the US is racing to protect from a trailing rival. It's multiple frontiers, feeding off each other, with Chinese models increasingly functioning as core infrastructure rather than distant competitors to benchmark against. Guarding your own lab's outputs matters less when your own industry keeps reaching across the fence to borrow the neighbor's tools.