China’s ambition to free itself from dependence on U.S.
tech giants in artificial intelligence is hitting a wall – not because of hardware shortages, but because of software. According to a new report from the South China Morning Post, local AI data centers are struggling to replace NVIDIA’s GPUs with Huawei’s alternatives, even under strict government mandates requiring at least 50% of chips in public projects to be domestically made.
The rule stems from Shanghai’s 2024 municipal guidelines, which this year became a national standard. It’s a direct push to curb reliance on foreign suppliers, particularly after years of tension over export bans and national security worries. When NVIDIA was allowed to ship its H20 GPUs to China, rumors swirled that the hardware carried hidden vulnerabilities or backdoors. NVIDIA denied the claims, but Beijing remained uneasy about depending so heavily on U.S. chips.
Huawei, together with SMIC, has stepped in to provide homegrown solutions. But the limitations are clear: SMIC can only manufacture on a 7nm process, since the advanced lithography tools needed for more cutting-edge designs are blocked by U.S. sanctions. This leaves Chinese AI hardware a step behind the industry leaders.
The bigger headache, however, is compatibility. NVIDIA’s GPUs are powered by the CUDA software ecosystem, a dominant standard for AI training. Huawei’s processors run on its CANN platform, which isn’t directly interchangeable. That means years of models, datasets, and infrastructure built on CUDA don’t easily migrate to Huawei hardware. According to insiders cited by SCMP, Huawei’s chips can handle running models, but retraining or developing entirely new ones is still far more efficient on NVIDIA’s GPUs.
As a result, cluster operators are stuck in a bind: the government demands domestic adoption, but the software divide forces painful compromises in efficiency, cost, and performance. The push for technological sovereignty is real, but the transition is proving messier than Beijing hoped.