DeepSeek is actively collaborating with Huawei to secure approximately 16,000 Huawei AI chips, aiming to resolve ecosystem issues for Chinese chips within a year by porting its proprietary compiler and TileLang-based environment to Huawei hardware. This strategy, articulated by Wenfeng, suggests that while Huawei's 950 SuperNode is about two years behind NVIDIA's GB200/GB300 and requires four cards to match one NVIDIA card, the primary bottleneck for China is production capacity, not the software ecosystem.
DeepSeek's move to port its software to Huawei chips directly challenges NVIDIA's CUDA moat, shifting the core constraint for Chinese AI compute from ecosystem to production capacity.