Alibaba T-Head

Zhenwu chips, AI processors, domestic GPU, spin-off IPO

CNprivateest. 2018t-head.cn (opens t-head.cn)Checked 11 Oct

The brief

Analyst view
  • Zhenwu 810E (January 2026) integrates training and inference, positioned between Nvidia A800 and H20, running at scale in Alibaba Cloud.
  • Zhenwu family achieved highest shipment volume among domestic Chinese GPUs in 2025; V900 unveiled September 2026.
  • Alibaba is spinning off T-Head as independent entity with separate IPO plans to accelerate growth beyond Alibaba Cloud.

Technical approach

As reported
Domestic design

Fully developed in China

Zhenwu is designed by Alibaba engineers, not modified to comply with export controls. This allows optimization for Alibaba Cloud's workloads.

Training and inference

Unified architecture

Zhenwu 810E integrates both training and inference, supporting language models, multimodal models and autonomous driving in one architecture.

Scale

Alibaba Cloud deployment

Zhenwu runs in clusters of tens of thousands on Alibaba Cloud, proven in production at scale for real-world validation.

Primer