China's AI compute buildout advances across hubs, chips and memory
Official capacity figures climbed again, while domestic GPU, memory and AI infrastructure suppliers reported new milestones.
• The National Data Administration said China's total AI computing power reached 2.45 million PFLOPS (FP16) by the end of July, with more than 85% of capacity in the eight national computing hubs and three computing-power/electricity-coordination zones and 1.45 million PFLOPS under the national monitoring and dispatch platform.
• Domestic GPU maker Biren Technology reported first-half revenue of 1.236 billion yuan, up 1,997.6% year on year, while gross profit rose to 527 million yuan and the net loss narrowed to 377.2 million yuan.
• ChangXin Technology said its in-house LPDDR6 chip has reached a peak rate of 12,800 Mbps and a maximum capacity of 16 GB, and Xiaomi said the Xiaomi 18 Fold will be the first device to ship with ChangXin LPDDR6 memory in September.
• China's first vehicle-mounted mobile token factory debuted with domestically developed computing power and an energy management system, using a 40-foot container design that integrates power, computing, cooling, storage and networking for rapid mobile deployment.
• CCTV reported that AI server growth is sharply increasing copper demand, with market estimates putting high-end copper foil for AI servers at about 24,000 tons in 2026, up roughly 260% year on year.
• Alibaba Cloud launched its Brazil region and said its global Qwen open-source models have surpassed 3 billion downloads, with more than 300,000 derivative models.
• OpenAtom Foundation released openKylin 3.0, moving the kernel to Linux 7.0 and adding an open base for intelligent agents plus multimodal features such as air gestures, voice input and screen reading.