🧠 Food for thought

Implications, context, and why it matters.

Huawei is leaning on system design, not leading-edge chips

• U.S. sanctions block Huawei from the most advanced chip manufacturing. Its Ascend 910C chips trail top-tier Nvidia models on single-chip performance and efficiency 1.
• Huawei counters with large clusters that connect hundreds or thousands of less advanced chips using proprietary high-speed interconnects 1.
• The China Mobile Guangdong contract follows that approach. The platform uses Ascend 910C chips plus Tiangong architecture and the Lingqu interconnect bus to build a large-scale computing platform for China Mobile Guangdong’s computing power service support platform, with a 160 exaFLOPS (FP16) target 2.
• This system-first method fits China’s push for technological self-reliance. It favors domestic control and supply chain control over raw power efficiency 1, 3.

China is building a national computing-power grid for AI

• The China Mobile Guangdong platform functions as one node in a state-led National Integrated Computing Power Network (NICPN) effort 3.
• The NICPN plan pulls together computing resources across regions, including domestic chip infrastructure and, where available, stockpiled U.S. chips, into a shared grid similar to an electrical network 3.
• The program treats compute as an interchangeable service that can be assigned on demand. Peng Cheng Laboratory (PCL, a Shenzhen-based research institute focused on advanced computing and AI) director Gao Wen compared the goal to making compute “like today’s electricity” 3.
• China Mobile Guangdong’s project excludes imported equipment, which helps build demand for domestic hardware. Separately, Huawei said it is expanding the Ascend software ecosystem through open-source work around CANN (Compute Architecture for Neural Networks, Huawei’s AI software stack for running models on Ascend chips) 2, 4.