Alibaba's T-Head subsidiary unveils Zhenwu V900 AI chip as part of expansion of its AI infrastructure stack and ambitious AI model pipeline.

T-Head, Alibaba's chip subsidiary, unveiled its next-generation Zhenwu V900 AI chip at the 2026 Apsara Conference in Hangzhou, marking a significant expansion of the company's AI infrastructure portfolio. The V900 is designed for both training and inference, delivering three times the performance of its predecessor, the Zhenwu M890, making it T-Head's most powerful in-house AI chip to date.

The V900 features 216GB of memory and offers 1,200GB/s of inter-chip bandwidth. It natively supports low-precision computing formats including FP8 and FP4, which improve efficiency for certain workloads and help reduce inference costs. The larger memory capacity reduces overhead associated with model partitioning and data movement for large-model training and inference.

Rather than operating as a standalone accelerator, the V900 is designed as part of a larger integrated system. T-Head uses its proprietary ICN Switch interconnect chips to connect multiple V900 chips into supernodes, providing native memory semantics and unified memory addressing while enabling full-bandwidth interconnection across thousands of AI chips. More than 1,000 V900 chips can work together as a single system, designed for workloads such as training trillion-parameter models and handling large-scale inference demand generated by AI agents.

At the conference, Alibaba showcased a new-generation supernode server integrating the Zhenwu V900, ICN Switch, Panmai smart NIC, and Zhenyue SSD controller, bringing computing, networking, and storage components together under a unified architecture. Combined with Alibaba Cloud's newly designed AI computing center network architecture, a single AI computing cluster using these chips can scale to as many as 500,000 accelerators.

Supernode servers based on the Zhenwu M890 have already entered large-scale commercial deployment and support models with more than two trillion parameters, including Qwen3.8 and Kimi K3. T-Head's Zhenwu series has served more than 650 enterprise customers across autonomous driving, finance, large language models, embodied AI, energy, and manufacturing sectors.

T-Head also detailed its Yitian server CPU roadmap. The Yitian 720 and Yitian 730 are scheduled to launch in the third quarter of 2027, with the 720 focusing on improvements in single-core performance, core density, and energy efficiency. The Yitian 730 will use a T-Head-developed CPU microarchitecture, delivering single-core SPECint2017/GHz performance up to 1.4 times that of the Yitian 710. The subsequent Yitian 750 will support T-Head's proprietary ICN inter-chip interconnect protocol, enabling the CPU to connect directly with Zhenwu AI chips. As AI servers grow more complex, tighter coordination between CPUs and AI accelerators is becoming part of system-level design.

The product roadmap reflects a broader full-stack chip strategy spanning CPUs, AI accelerators, networking, and storage. As model sizes and inference workloads grow, bottlenecks increasingly extend beyond raw compute to chip-to-chip communication, memory capacity, network bandwidth, and data access.

The Zhenwu V900 is expected to enter mass production in the first quarter of 2027. Alibaba Group CEO Wu Yongming stated during the conference that T-Head expects its annual AI chip shipments to increase as its products see wider adoption. T-Head also showcased its full-stack chip solutions and held hands-on workshops around the T-Head SAIL software stack, covering model training, inference, and software optimization.

For China's AI chip industry, T-Head's strategy highlights a growing focus on system-level infrastructure as domestic chipmakers look to compete not only on accelerator performance but also on the interconnect, memory, and software layers needed to run increasingly large AI models.