Domestic AI compute demand and hardware deployment stay in focus
Chinese companies highlighted domestic-compute-trained models, larger AI clusters, agentic devices and continued strength in semiconductor-related stocks.
• iFlytek launched Spark-Audio-1.0-Preview, a speech foundation model trained entirely on domestically sourced computing power and designed to process speech, tone, emotion and background sounds end to end.
• A Huawei executive said supernode clusters with more than 100,000 cards could become standard in 2027, while 10-trillion-parameter models may soon become the baseline for ultra-large models; Huawei also released what it called the world’s first 3D data center.
• Chinese semiconductor shares and ETFs rose as AI-computing demand and chip-price increases supported the sector, with Cambricon up 5.51% and the Huabao Electronics ETF reporting more than 40 billion yuan in net inflows.
• Nubia launched the NaviX Ultra, the second-generation Doubao phone, with Doubao Phone Assistant integrated at the system level; Nubia said launch sales revenue exceeded 100 million yuan in one second, with subsidized prices starting at 5,499 yuan.
• vivo introduced four BlueLM models and a system-level BlueLM Harness, while its Xiao V Pro assistant can reportedly call more than 6,000 atomic capabilities and support over 10,000 task types.
• vivo also unveiled Xiao V Cowork for starting and managing PC-side agent tasks from a phone and said it is researching a 30-billion-parameter BlueLM MoE model for on-device deployment.
• Jiemian Caixin said its Kuobian system will combine exclusive information collection, a multimodal financial database and a finance-focused large language model as part of an effort to build self-reliant and controllable financial-information infrastructure.