Chinese semiconductor firm Sophgo has adapted its SC11 FP300 compute card to support DeepSeek’s reasoning model.

This adaptation underscores China’s efforts to develop indigenous AI systems amid US export controls limiting access to foreign chips.

The performance of Sophgo’s compute card was verified in tests by the China Telecommunication Technology Labs (CTTL), a research institution affiliated with the Ministry of Industry and Information Technology.

The results showed stable execution of DeepSeek’s R1 model, according to a statement from Sophgo.

🔗 Source: South China Morning Post

🧠 Food for thought

1️⃣ China’s seven-year AI sovereignty strategy accelerates under pressure

The Sophgo-DeepSeek collaboration is not a reactive measure but part of China’s deliberate long-term AI strategy dating back to 2017, when Beijing launched its “New Generation Artificial Intelligence Development Plan” targeting a $150 billion AI industry by 2030 1.

This latest development directly addresses a vulnerability in China’s AI stack that government planners identified years ago: dependency on foreign-made chips for advanced computing.

As U.S. export controls tightened, Chinese AI companies have increased investments in domestic alternatives—with iFlyTek notably claiming to be the only Chinese AI developer fully training its models with domestic chips like Huawei’s Ascend 910B 2.

The verification of Sophgo’s compute card by a government-affiliated lab highlights China’s “top-down approach” to AI development, which emphasizes systematic collaboration between government agencies, private companies, and research institutions 3.

These coordinated efforts to build a complete domestic AI supply chain explain why China’s semiconductor sector continues to show strong growth despite export restrictions, with companies rapidly adapting their business models to new constraints 4.

2️⃣ Performance sacrifices reveal strategic priorities in China’s tech sector

Chinese AI developers are making calculated trade-offs between optimal performance and technological independence, with multiple companies explicitly accepting development delays to reduce dependency risks.

iFlyTek’s chairman publicly acknowledged that using domestic chips extends AI model development time by approximately three months compared to using Nvidia’s chips—a significant but apparently acceptable cost 2.

The efficiency gap between domestic and foreign chips remains substantial but is narrowing, with iFlyTek reporting that Huawei’s chips improved from 25% to 73% efficiency compared to Nvidia A800 chips within just six months 2.