China’s AI infrastructure push expands as DeepSeek prepares new model and IPO
Chinese firms announced a planned 100,000-card AI cluster, advanced-packaging production and new model and financing initiatives, while domestic demand for inference infrastructure remains strong.
• JD Cloud and Moore Threads plan to build a 100,000-card domestic AI data-center cluster, following a 10,000-card cluster developed with Moore Threads and other partners.
• DeepSeek has hired CITIC Securities to prepare a STAR Market IPO, targeting an application this year and a listing next year, with proceeds intended for compute infrastructure, model development, in-house chip research and talent incentives.
• DeepSeek plans to release V4.1 Flash around September 10 after internal and external tests reportedly found it ahead of V4 Pro on performance, cost, speed and total response time.
• Ant Group open-sourced Ling-3.0-flash-VL, its first native multimodal model in the Ling series, with 124 billion total parameters, 5.5 billion activated parameters and a 256K-token context window.
• Huahai Qingke said its domestically developed Master-P510APEX fully automated board-level CMP system had entered a mass-production line, marking a transition from development to commercial application for the equipment.
• SK hynix said it has not decided whether to sell part of its stake in its Chongqing packaging plant, while reviewing measures to improve the competitiveness of its packaging business.
• Huawei’s Pura X View and Mate XT 2 Ultimate Design reportedly support downloading and running 6GB multimodal-enhanced or 15GB multimodal mixture-of-experts models locally without an internet connection.
• China’s Ministry of Commerce rejected US allegations of industrial-scale model distillation by Chinese AI companies, calling the claims unsupported and opposing the related advisory.