Author: Evolving AI - Bewertung: 0x - Views:0
China just stopped playing Nvidia’s game and started building its own. For years, Chinese tech giants depended on cut-down H800, A800, and H20 GPUs, but in 2025, Beijing quietly flipped the tables: state media called H20 “unsafe,” regulators hauled Nvidia in, and companies were pushed to cancel new GPU orders. In this video, we break down how Huawei, Alibaba, Baidu, and Cambricon are now rolling out their own AI chips at cluster scale, not as experiments, but as the backbone for DeepSeek-class frontier models and national cloud infrastructure. We start with Huawei’s Ascend 910B and dual-chiplet 910C powering Atlas 900/950 SuperPoD systems that scale to thousands of chips and exaflops of FP8, all programmed through MindSpore and CANN as China’s answer to PyTorch and CUDA. Then we move to Alibaba’s Hanguang 800 and PPU accelerators running inside a 22,000-chip China Unicom cluster, Baidu’s Kunlun 3 P800 powering a 30,000-chip AI supercluster and Qianfan-VL multimodal models, and Cambricon’s comeback with its MLU 590 and upcoming MLU 690 pushing toward H100-class performance. Together, these aren’t just “alternatives” anymore – they’re the foundation of a domestic AI hardware stack built on DUV, liquid-cooled racks and petabytes-per-second interconnects, designed to break Nvidia’s leverage over China’s AI future. And if you want the real story behind the world’s fastest-moving tech and AI breakthroughs, make sure to like and subscribe to Evolving AI for daily coverage.
SOCIAL SHARE CARD GENERATOR