"US export controls primarily target the digital loop, restricting access to advanced chips used for frontier model training - but are not well suited to addressing the physical loop of deployment-driven data creation and accumulation across China's manufacturing base."
"As open models reduce the compute required for effective deployment, China's ability to generate proprietary industrial data at pace and scale becomes increasingly independent of access to cutting-edge hardware."
"Alibaba's Qwen model family is now the largest ecosystem on Hugging Face with over 100,000 derivative models, with seven of the ten most downloaded models coming from Chinese labs."
"Roughly 80% of US startups use Chinese base models to build derivative applications, indicating deep commercial penetration into US enterprises."
China's open-source AI strategy is establishing a competitive edge that US export controls do not effectively counter. The US-China Economic and Security Review Commission highlights that these controls focus on digital loops, limiting access to advanced chips, but overlook the physical loop of data generation in China's manufacturing. Chinese labs are leveraging permissive licensing and pricing to build a global diffusion engine, leading to significant industrial data advantages. The adoption of Chinese models is widespread, with a large percentage of US startups utilizing them for derivative applications.
Read at Computerworld
Unable to calculate read time
Collection
[
|
...
]