Moonshot taps Alibaba Cloud as China model and compute buildout advances
China’s AI compute news centered on Moonshot’s reported GPU lease, large data-center campus plans, domestic model activity, and tightening chip and memory supply-chain signals.
• Moonshot AI reportedly leased 20,000 Nvidia GPUs from Alibaba Cloud, making up a significant proportion of compute capacity for its Kimi models.
• RedNote could build a 600MW data center campus in Inner Mongolia, while DeepSeek is also reportedly considering a 1GW campus in Ulanqab.
• DeepSeek said its V4 Flash API performance degradation on Aug. 4 had been resolved, after users and OpenCode reported capacity shortages and errors during heavy traffic.
• OpenRouter said DeepSeek-V4-Flash ranked first globally last week with 7.22 trillion tokens in usage.
• Alibaba’s Qwen3.8 model has been officially released and is expected to be open-sourced next week, while Alibaba’s Qianwen Office has begun public beta testing.
• MiniMax released the core weights of its H3 omni-modal video model, and Insta360 is reportedly integrating Alibaba’s Qwen model into mainland China versions of its AI hardware.
• Feilong Auto Components said current liquid-cooling orders remain in the small-batch stage and have not reached its separate disclosure threshold.
• HP, Asus and Acer have reportedly begun limited use of CXMT DRAM chips in laptops, while CXMT is prioritizing most of its capacity for Chinese customers.