China has reportedly issued a sweeping ban on foreign AI chips in any data center backed by government money, according to a report by Reuters. The move applies retroactively to builds still in early stages, meaning accelerators from Nvidia, AMD, and Intel may need to be removed or replaced if already installed.
The report, citing unnamed sources with direct knowledge, claims regulators have instructed state-funded projects to use only Chinese-made silicon going forward. That includes domestic chips from Huawei, Cambricon, and Enflame, among others.
While the official directive has not been published, the guidance would mark a formal shift from earlier policies that merely discouraged foreign chip purchases. Under the new dictat, even parts like Nvidia’s H20 — a model specifically designed to comply with U.S. export controls — will now be off the table. This follows Chinese port crackdowns on all Nvidia imports in early October.
Projects that are less than 30% complete are allegedly being told to abandon or remove any foreign chips already in use. For Nvidia, which once held more than 90% of China’s AI accelerator market, that effectively kills prospects for a comeback via more custom silicon. Last month, Nvidia CEO Jensen Huang said that the company’s Chinese market share has effectively plummeted to zero.
The guidance appears to target not only the H20, but also Nvidia’s higher-end H200 and B200, which are restricted under U.S. export rules but have still made their way into Chinese data centers through unofficial channels. That flow of gray-market parts may now face stricter internal scrutiny, with implications for training clusters already under construction.
Over the past two years, China has funneled more than $100 billion into AI infrastructure projects, most of them aligned with provincial or national goals for data sovereignty. That public investment is what gives Beijing’s directive real teeth. But while Chinese chipmakers have made strides, they continue to trail Nvidia and AMD in software tooling and performance density. Huawei’s Ascend line, often cited as the most mature domestic option, still lacks full parity with CUDA-based stacks.
Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.
