Huawei and DeepSeek spotlight China’s push across AI chips, packaging and models
China’s AI compute sector saw developments spanning domestic chip adoption, advanced packaging, memory supply, and open model infrastructure.
• Huawei executive Eric Xu said sales of homegrown AI chips have surpassed Nvidia’s in China, arguing that domestic chips may be less advanced but are less exposed to sudden export bans.
• DeepSeek open-sourced its TileLang compiler, compute libraries and distributed communication libraries for Huawei’s Ascend platform, developed with Huawei toward a 128-card Ascend 950 supernode.
• Hubei Xingchen Technology brought China’s first 3D advanced-packaging mass-production line online in Wuhan’s Optics Valley, following combined investment of more than 7 billion yuan across its two project phases.
• TrendForce forecast Q4 contract-price increases of 10%-15% for conventional DRAM and 15%-20% for DRAM including HBM and for NAND flash, citing continued allocation of advanced capacity to high-performance servers and accelerating AI-related demand.
• Bilibili’s Index-Translate family released 2B, 9B and 35B-A3B preview text models on Hugging Face and ModelScope, with support for 150 languages and translation instructions covering terminology, formatting and content preservation.
• Baseten added Moonshot AI’s Kimi K3 and GLM 5.3 to OpenAI’s enterprise marketplace, making Kimi K3 the first Chinese open-source model reported to enter OpenAI’s enterprise procurement system.
• Dongxin Semiconductor said its compute-in-memory and interconnect chip project has an estimated 42-month development timeline and will use advanced packaging to integrate memory, interconnect and computing units on one chip.
• Doubao expanded its AI assistant with flight and train bookings, ride-hailing and navigation, while a separate personal-assistant product reportedly named Xiaodou remains in internal testing.