Domestic chips and compute networks take focus as China AI model makers scale
China’s AI compute news centered on domestic chip adoption, new model releases, infrastructure planning and semiconductor supply-chain investment.
• MiniMax said M3 Pro is expected to reach about 3 trillion parameters, while it is adapting M3 and H3 to domestic chips and preparing a large domestic computing cluster to take over real production traffic.
• Zhipu launched and open-sourced GLM-5.3-Flash, a 320B-parameter native multimodal model with 18B active parameters, and said its pre-release Ox-Alpha traffic was powered by domestic chips.
• Alibaba’s Tongyi Qianwen team released Qwen3.8-Flash as open source, a multimodal MoE model based on the next-generation Qwen4 architecture with 125B main parameters and 6B activated per token.
• MIIT officials said China has developed nearly 200 key AI standards and will speed up planning and construction of new-generation communication and computing power networks while strengthening 6G R&D.
• Tencent executive Dowson Tong said insufficient computing power has slowed Hunyuan large-model training, iteration and product rollout, while acknowledging outside criticism that Tencent has been slow on AI.
• Huawei reportedly bid for an Egyptian government AI data center project, proposing at least 1,408 Ascend 950-series chips for AI training and a 12-month local infrastructure buildout.
• Changxin Technology fully exercised its IPO over-allotment option, issuing an additional 1.003 billion shares and raising net proceeds of 66.310 billion yuan for memory wafer production-line upgrades.
• Samsung Electronics and SK Hynix are reportedly continuing investments in NAND wafer fabs in China, including Samsung process upgrades at its Xi’an X2 fab toward ninth-generation V-NAND.