Beijing Launches Subsidy Policy for Domestic GPU Procurement
The Beijing municipal government offers subsidies to enterprises purchasing domestically developed GPUs for intelligent computing centers, with a maximum subsidy of ¥50 million per company. Technical requirements include a minimum 60% localization rate to boost domestic chip adoption.
China’s Big Three Telecom Operators Plan $70B Investment in Computing Centers
China Mobile, China Unicom, and China Telecom prioritize domestic server deployment (over 80% localization) in regions like the Yangtze River Delta and Ningxia, while optimizing resource allocation under the "East Data, West Computing" initiative.
National Integrated Computing Network Accelerates
China’s National Data Bureau aims to establish a nationwide computing collaboration system by 2025, addressing regional imbalances and enhancing domestic GPU cluster efficiency.
Huawei’s Ascend 910C Matches NVIDIA A100 Performance
The Ascend 910C achieves 82% of the A100’s performance in tests. Its adoption rate in Ulanqab Computing Center surged from 10% to 58%, signaling strong import substitution.
Moore Threads Deploys 1,000-Card Clusters in Beijing and Nanjing
The company’s "Kuā É" cluster solution supports trillion-parameter AI model training, improves computing efficiency by 50%, and maintains CUDA compatibility.
Hygon, Cambricon See Surging GPU Shipments
Hygon’s Q2 DCU chip shipments rose 217% YoY. Cambricon’s Siyuan 590 secured orders from ByteDance, while Sugon adapted over 2,000 domestic software tools.
U.S. Tightens AI Chip Export Controls
The Biden administration adds China to its AI technology embargo list, restricting NVIDIA and others from exporting advanced GPUs, spurring faster domestic substitution.
NVIDIA Unveils Blackwell Ultra and Rubin Architecture Chips
The new Blackwell Ultra GPUs, 14x faster than predecessors, drive surging demand for optical modules. Microsoft and Meta preordered 3.6 million units.
DeepSeek Open-Sources GPU Efficiency Breakthrough
Its FlashMLA decoding kernel optimizes Hopper GPU performance, achieving 580 TFLOPS on H800 and cutting large-model inference costs.
Liquid Cooling Demand Soars
ByteDance and Tencent report 5x higher costs for data center retrofits. Telecom operators’ liquid cooling purchases surged 700% YoY, creating a $10B+ market.
Summary
This week highlighted policy-driven growth, technological leaps, and intense global competition in GPU computing. China’s subsidies and cluster projects accelerated domestic chip adoption, while U.S. export controls and NVIDIA’s innovations underscored the strategic value of computing autonomy. Supporting technologies like liquid cooling and optical modules also saw explosive demand.