China’s AI compute pipeline expands from memory fabs to local AI PCs
Chinese companies announced major memory infrastructure spending, advanced-model updates, compute-service agreements and new systems for on-device AI workloads.
• ChangXin Technology plans 24.1 billion yuan for a technology R&D project and 10.8 billion yuan for a phase-two DRAM wafer back-end testing base, funded partly through 13 billion yuan in excess proceeds and a 5 billion yuan investment in its memory subsidiary.
• Xiaomi promoted large-model team head Luo Fuli to level 22, the company’s highest job grade, after releasing and open-sourcing the MiMo-V2.6 Pro and Flash native omni-modal models on Sept. 22.
• Moonshot AI’s API registry exposed a callable “kimi-k3-1” identifier, while an official teaser pointed to a forthcoming K3.1 or K3 11111 model with a potential 1-million-token context window and multiple reasoning modes.
• Fengzhushou signed a 5.717 billion yuan computing-power services agreement and a 3.96 billion yuan computing-server procurement agreement.
• Alibaba’s Qianwen app added authorized access to Quark Netdisk, allowing users to search, organize and read cloud files in conversations and invoke Quark Agent capabilities for tasks based on stored materials.
• ACEMAGIC and GMKtec introduced desktop AI systems based on AMD’s Ryzen AI Max+ PRO 495, with configurations reaching 192GB of memory; GMKtec said its EVO-X5 Pro can run 320-billion-parameter models locally.
• NVIDIA released the Open Agent Safety Platform, including OpenShell software designed to set boundaries for AI agents running on CPUs and support governance across agent hardware and robotic systems.