China expands AI capacity, domestic stacks and model offerings
China’s AI compute sector saw new data-center capacity, policy support, domestic hardware-software integration and model releases, while power and operating costs remain key infrastructure issues.
• The Ulanqab Vision Galaxy Base began operations in August with 120,000 square meters of space, direct renewable-power connections, parallel capacity for 1 million chips and petaflop-scale computing, according to CCTV Finance.
• Guangxi’s 15th Five-Year Plan calls for an AI-focused fund cluster with at least 50 billion yuan in capital, along with products including AI loans and computing-power loans.
• China Telecom open-sourced Xing4.0-29B-A4B, a 29-billion-parameter model with 4 billion activated parameters, a native 256K context window expandable to 512K, and a fully domestic training and deployment stack, the company said.
• Huawei said its Ascend ecosystem has reached an inflection point, citing more than 5,200 average monthly active users in the CANN community, over 30,000 lines of code merged daily and more than 40 models pretrained on Ascend.
• Alibaba released Qwen3.8-LiveTranslate, a simultaneous-interpretation model supporting 60 languages, speaker separation, bilingual source-and-translation display and a reduction in average lagging latency from 2.8 seconds to 2.3 seconds.
• DeepSeek said weekends designated as make-up workdays and China’s statutory public holidays will be charged at off-peak API rates.