China’s race to cut dependence on U.S.
technology has entered a new phase, with Beijing pushing local firms to embrace homegrown AI chips in their datacenters. This move comes as trade tensions deepen and Washington grows wary of advanced semiconductor exports to China.
The spotlight now falls on Cambricon, one of China’s most prominent AI chip startups, which plans to raise roughly $560 million (about 4 billion yuan) to expand its capabilities. The company aims to provide alternatives to NVIDIA and AMD, developing processors that can power large-scale AI models and cloud services. Its Siyuan series already targets datacenter and cloud computing needs, and more advanced chips are being prepared for training next-generation large language models.
Beijing is setting clear requirements: more than half of the AI chips in new data centers must be sourced from domestic vendors. This policy strongly favors companies like Cambricon and Huawei, which is aggressively promoting its Ascend lineup. Huawei’s Ascend 910C, in particular, is rumored to outperform NVIDIA’s H100 in training benchmarks, while its CloudMatrix 384 rack-scale system is being positioned as a competitor to NVIDIA’s Blackwell NVL72.
Still, challenges remain. Despite the hardware push, China lacks a mature equivalent to NVIDIA’s CUDA ecosystem, making it difficult for developers to fully harness local chips. Performance gaps are also evident, with reports suggesting that DeepSeek’s R2 AI model was delayed because current domestic hardware fell short in training efficiency.
For now, Chinese AI companies must straddle both worlds – leaning on U.S. technology for cutting-edge results while investing heavily in local solutions. Whether Cambricon’s fundraising drive will accelerate breakthroughs remains to be seen, but the momentum clearly signals Beijing’s determination to shift the balance of power in global AI hardware.