Huawei Launches CloudMatrix 384 Ultra-Node Architecture
Huawei Cloud released an AI computing cluster with 384 Ascend NPUs + 192 Kunpeng CPUs in full peer-to-peer interconnect. Single-card inference throughput reaches 2,300 Tokens/s, supporting 160,000-card cascading. Features innovative "daytime inference/nighttime training" scheduling, deployed at 1,300+ clients including iFLYTEK and CAS.
Pangu 5.5 Model Breakthrough on Domestic Compute
Huawei unveiled its 718B-parameter MoE model (Pangu 5.5) fully trained on Ascend hardware. Achieves world-class knowledge reasoning/math capabilities, proving China's ability to train top-tier models domestically.
Aofei Zhisuan's GPU Scheduling Patent
Chengdu Aofei Zhisuan filed a patent for dynamic GPU scheduling using thermal-expansion monitoring and workload migration, optimizing resource allocation to resolve overheating bottlenecks.
Domestic GPU Commercial Inflection Point
CITIC Securities reports China's GPU market will hit ¥120B by 2025. Domestic chips (e.g., Cambricon, Hygon) now cost less than NVIDIA A100 while delivering 90% of performance.
Dongshan Precision Acquires Optical Module Maker Source Photonics
¥5.9B acquisition targets Source Photonics (global #4 in 800G optical transceiver shipments) to strengthen domestic computing supply chains.
AMD Debuts MI350 GPU & Hyper-Node Plan
CDNA4-based MI350 GPUs feature 288GB HBM3e, delivering 35x Llama inference gains (FP4 optimized). "Helios" hyper-node cabinets (2026) will support 1,024 GPU interconnects.
TSMC Arizona Begins 4nm AI Chip Production
First 4nm wafers (including NVIDIA Blackwell/Apple chips) completed at Arizona fab. CoWoS packaging still Taiwan-dependent; US packaging plant to open by 2026.
NVIDIA Accelerates European Sovereign AI
GTC Paris announced 3,000 EFLOPS Blackwell deployments in France/Italy/UK. Partners with Siemens for industrial AI cloud (digital twins/autonomous driving).
Jensen Huang Acknowledges Huawei's Ultra-Node Lead
NVIDIA CEO publicly stated Huawei’s CloudMatrix 384 "has surpassed NVIDIA," shifting competition focus to cluster efficiency/ecosystems over chip process.
AWS & SK Build Korea's Largest AI Center
$4.5B project in Ulsan deploys 60,000 high-end GPUs (103MW capacity). Operations start 2029, targeting Asian AI demand.
Summary
This week spotlighted ultra-node architecture competition (Huawei/AMD/NVIDIA) and China's substitution breakthroughs (model training, optical modules, cost advantages). Advanced packaging (TSMC) and edge computing (European industrial cloud) emerged as new growth vectors.