Hot Take: US says China stole its AI, here's what they won't tell you

In late July, US officials accused China's Moonshot of "industrial-scale distillation" of Anthropic's AI and threatened sanctions.

So what exactly is "distillation"? Does it amount to copying? And how did China's AI models achieve their rapid breakthroughs?

Distillation is standard practice, widely used by US tech giants like Google and xAI.

Calling it "theft" is seen by many as a commercial tactic to slow down competitors, not a technical reality.

US officials charge that large-scale unauthorized distillation violates terms of service and raises national security concerns. 

China's Commerce Ministry has hit back, accusing Washington of "double standards" and "typical AI hegemony." It pointed out that many US AI companies have distilled Chinese models in their own R&D, using the very technique the administration is now trying to ban.

Elon Musk's team used it for Grok. Google even sells a distillation service. And Jensen Huang, Nvidia's CEO, put it best: "Distillation, learning from AI, learning from other sources of knowledge, is fundamental to intelligence." 

That's not theft – that's how intelligence works.

The key point is this: distillation alone cannot produce a model with the performance of Kimi K3. Moonshot says the leap came from its own architecture, a new optimizer called Moon Clip, a linear-attention system and attention residuals – breakthroughs Musk once praised. Original engineering, not a copy, at about a third of Anthropic's flagship price per task.

And here's a telling contrast: on OpenRouter, US companies now run nearly 60% of their tokens on Chinese models. In July, Hugging Face – an American platform – had to call in a Chinese model to fight off an attack that US frontier models couldn't handle. Washington wants to cut off what Silicon Valley is already buying.

Distillation was never the whole story. It's about who gets to define innovation. And China's advancements have made that conversation truly global.