China is requiring publicly owned data centers to use at least 50% domestically produced chips, as Beijing moves to decrease reliance on foreign technology amid tightened US export controls.

This policy, initially proposed in Shanghai, has now been applied nationwide, according to sources familiar with the matter.

The guidelines aim to support China’s semiconductor industry and strengthen AI infrastructure, as the US continues to restrict exports of advanced chips like Nvidia’s H100 and H800.

Over 500 new data centre projects were announced in China in 2023 and 2024, according to MIT Technology Review, citing KZ Consulting.

Chinese chips are reported to be sufficient for running trained AI models, but Nvidia chips remain preferred for training new models.

🔗 Source: South China Morning Post

🧠 Food for thought

1️⃣ The 50% chip mandate accelerates China’s decade-long self-sufficiency timeline

This nationwide requirement represents a significant acceleration of China’s semiconductor independence strategy that has been building for nearly a decade.

Since 2014, China has invested over $40 billion toward achieving 70% chip sufficiency by 2030, making this data center mandate a concrete policy mechanism to drive domestic adoption1.

The timing is particularly notable given China’s “Made in China 2025” initiative, launched around 2015, which set ambitious targets including $305 billion in semiconductor output by 2030—up from just $65 billion in 2016 when China captured only 33% of its domestic market2.

By requiring immediate 50% domestic sourcing in government data centers, Beijing is creating guaranteed demand for Chinese chipmakers while they work toward broader technological parity.

This approach focuses on how China currently achieves self-sufficiency in mature 22/28nm processes and aims for a 40% share of the mature process market by 2030, emphasizing areas where domestic capabilities can compete effectively3.

2️⃣ Software ecosystem incompatibilities create unexpected technical hurdles for mixed-chip deployments

The mandate reveals a complex technical challenge that goes beyond simply swapping hardware—different chip manufacturers use incompatible software platforms that require significant adaptation work.

AI chips typically run on proprietary software ecosystems, with Nvidia using CUDA and Huawei using CANN, meaning models developed on one platform need extensive modification to run on another manufacturer’s chips.