T-Head unveils new AI chip, expands server CPU roadmap
Alibaba's chip unit, T-Head, has introduced its Zhenwu V900 AI chip for training and inference. The company also detailed its roadmap for future Yitian server CPUs.
Alibaba's chip subsidiary, T-Head, unveiled its next-generation Zhenwu V900 AI chip at the 2026 Apsara Conference in Hangzhou. The chip is designed for both AI model training and inference workloads.
The announcement signifies T-Head's move toward a full-stack chip strategy, encompassing computing, storage, and networking, as AI models grow and agentic applications become more prevalent.
The Zhenwu V900 is T-Head's most powerful in-house AI chip to date, reportedly offering three times the performance of its predecessor. It features 216GB of memory and supports low-precision formats like FP8 and FP4, potentially improving efficiency and reducing inference costs.
Designed for system-level integration, the V900 utilizes T-Head's ICN Switch interconnect chips to connect multiple units into supernodes. This configuration allows for unified memory addressing and full-bandwidth interconnection across thousands of AI chips, enabling a single system of over 1,000 V900 chips for training trillion-parameter models.
Alibaba also showcased a new supernode server integrating the V900, ICN Switch, Panmai smart NIC, and Zhenyue SSD controller. This system, combined with Alibaba Cloud's AI computing network architecture, can scale to 500,000 accelerators.
Additionally, T-Head provided an update on its Yitian server CPU roadmap, with the Yitian 720 and 730 slated for Q3 2027. A future Yitian 750 will support T-Head's ICN protocol for direct CPU-AI chip communication. The Zhenwu V900 is expected to enter mass production in Q1 2027.