Short on compute, AI organizations are busy finding every possible loophole. China stands at the center of this conversation.
China-based AI developers have finally left the throes of academic experiments, moving to commercial-scale model distillation. And they’re leveraging Silicon Valley’s AI models for the same purpose.
At least this is what Reuters’ research tells us.
Chinese firms have automated the collation of all the outputs, with specialists curating and scoring responses to create training datasets. Because the end goal for them is efficiency.
While they care about developing their own domestic models, China-based firms find it easier to use US AI models for the foundational data. They’re aware that one expensive query to any frontier Silicon Valley model produces high-value examples that will then be reused to train cheaper, domestic models.
As per Reuters, the developers collected 5,000 answers from OpenAI’s o1 model, translated them through GPT-4o mini, and then collated over 11,000 bilingual examples for training Alibaba’s Qwen2.5-72B model.
And then following the same was Alibaba researchers leveraging GPT-4 and GPT-4o for smaller Qwen models.
But there’s a massive hurdle here.
Where do Chinese researchers land?
Over-relying on the US’ AI models could stifle original innovation. This leaves domestic developers dependent on breakthroughs they’ll never own.
However, the government isn’t one to fear staying behind.
It hasn’t explicitly ordered the distillation of American models, while state policies do encourage it, especially to lower computing requirements and make AI affordable across China’s manufacturing sector and other industries.
Even with backing from national authorities, developers are worried about the gray area. Where’s the beginning and end of the use of foreign models? Is it limited to training data?
Diplomatic clashes are always expected.
AI has intensified the tensions between the US and China. And it isn’t as if the US isn’t aware of China’s practices. In July 2026, in Chengdu, officials from the US Commerce Department issued a warning that concerns IP theft. Meanwhile, Chinese officials defended the distillation process, saying it’s a standard ML procedure.
US will never be ready to accept the AI performance gap. To curb China’s reach into AI models, the administration has issued new export controls targeting its remote access to advanced chips through overseas data centers in Singapore, Malaysia, and more.
Sounds like damage control after Huang’s affairs left a dent in a relationship already brimming with friction. Not much is going to change- irrespective of the undercover export of chips to China, the US will do everything in its power to restrain their access to the chips.
And this is merely the beginning.


