Highlights

• Alibaba Cloud held its 2026 Apsara Conference on 22-24 Sep 26 and framed AI as a long-term scaling opportunity spanning models, cloud infrastructure and proprietary compute. Key takeaways include: a) continued scaling of its model roadmap toward 5-10T parameters, alongside new multimodal model launches; b) further expansion of its proprietary compute stack with the Zhenwu V900, Yitian CPU roadmap and Panjiu supernode; and c) expansion of global DC capacity to 20GW by 2032 (10x vs 2022), and accelerating the deployment of proprietary chips. Maintain BUY with an unchanged target price of HK$188.00 (US$192.00).

Analysis

• Model roadmap advances toward 5-10T parameters, with RSI extending into training and inference optimisation. Alibaba Group (Alibaba) highlighted continued progress in large models, with Qwen4 currently under training and future Qwen4.5 and Qwen5 versions expected to scale to 5-10T parameters. Meanwhile, Recursive Self-Improvement (RSI) has started to extend across model training and inference optimisation. Qwen3.8-Max autonomously built training workflows and generated training data, completing 33 effective iterations with virtually no human participation, while also optimising the Qwen3.8-Flash inference framework to improve single-instance throughput by 96%. Separately, Alibaba broadened its multimodal portfolio with Qwen-Image-3.1, Happy Shrimp 1.1, HappyOyster 2.0 Preview, Qwen-Audio-3.1 and Qwen3.8-LiveTranslate, while its next-generation video-generation model is scheduled for launch in November.
   
• Proprietary chip roadmap expands from individual processors toward integrated AI supernodes. The M890 supernode has already supported inference for 2T-parameter models. Alibaba unveiled its next-generation Zhenwu V900 AI accelerator, delivering around 3x the performance of the M890, with 216GB HBM and 1,200GB/s inter-chip bandwidth, and is expected to enter mass production in 1Q27. T-Head also outlined its Yitian CPU roadmap, with Yitian 720/730 planned for 3Q27 and the subsequent Yitian 750 designed to support Alibaba's proprietary ICN interconnect, enabling tighter CPU-to-AI-accelerator integration. At the system level, Alibaba unveiled its next-generation Panjiu supernode server, integrating Zhenwu V900, ICN Switch, Panmai smart NIC and Zhenyue SSD controller chips with Alibaba Cloud's latest intelligent computing-centre network architecture to enable system-level coordination across compute, storage and networking. A single AI cluster can scale to up to 500,000 accelerators, with the architecture designed to evolve toward wide-area supernodes. The roadmap highlights Alibaba's shift from standalone chip development toward full-system AI infrastructure optimisation.