On September 22, the 2026 Yunqi Conference hosted by Alibaba Group opened at the Hangzhou International Expo Center. During the event, Alibaba's T-Head unveiled its next-generation unified training-and-inference AI chip, the Zhenwu V900, and outlined plans for a super-node server and future chip products.

Zhenwu V900 Triples the Performance of the M890

According to the company, the Zhenwu V900 uses T-Head's in-house parallel computing architecture and delivers three times the performance of the previous-generation Zhenwu M890, making it suitable for training and inference of trillion-parameter large models.

The chip carries 216GB of high-capacity memory and a chip-to-chip interconnect bandwidth of 1,200GB/s, with native support for FP8 and FP4 low-precision computing. By supporting computing modes across different precision levels, the Zhenwu V900 can handle a range of tasks, from high-precision model training to low-precision inference and ultra-low-precision inference.

In large-model inference scenarios, low-precision computing such as FP8 and FP4 can reduce data transfer volume and computing costs, boosting computing power density while preserving model performance. Alibaba said the Zhenwu V900 will focus on large-parameter models and high-concurrency AI applications.

The Zhenwu V900 is slated for mass production and sale in the first quarter of 2027, with gradual large-scale deployment across Alibaba Cloud data centers.

ICN Switch Enables High-Speed Interconnect Across Thousands of Chips

To meet the memory-capacity and chip-interconnect demands of trillion-parameter models, the Zhenwu V900 does not operate as a standalone single chip. Instead, it works alongside T-Head's in-house networking, interconnect, and storage chips to form a computing power system.

Among these, the ICN Switch interconnect chip supports full-bandwidth, high-speed interconnection at the scale of thousands of chips. Connected through the ICN Switch, multiple Zhenwu V900 chips can achieve native memory semantics and unified memory addressing, allowing thousands of chips to work together like a single "super chip."

This architecture reduces the data-movement overhead in large-model training and inference and improves communication efficiency between chips. For ultra-large-scale models that frequently exchange parameters and intermediate results, chip-to-chip interconnect capability has become a major factor in overall performance.

Panjiu Super-Node Server to Launch in 2027

At the server level, Alibaba also unveiled a new Panjiu super-node server built around the Zhenwu V900.

The server integrates the Zhenwu V900, the ICN Switch, the Panmai smart network card, and the Zhenyue SSD controller chip, covering computing, interconnect, networking, and storage. Combined with Alibaba Cloud's next-generation network architecture, it forms a computing power system with coordinated hardware and software.

According to information released by Alibaba, a single AI cluster can scale to 500,000 chips, providing computing power for large-scale model training and inference tasks. The Panjiu super-node server is scheduled to launch in the first quarter of 2027.

What Alibaba unveiled this time is not a single chip product but a complete product portfolio spanning AI chips, interconnect chips, smart network cards, storage controllers, full server systems, and the cloud platform.

Zhenwu M890 Super-Node Already in Service

In addition to the next-generation Zhenwu V900, Alibaba Cloud's existing Lingjun Zhenwu M890 super-node instance, the GP9A, is already available to customers.

The super-node has reportedly run large models with more than 2 trillion parameters, making it one of the earliest super-node computing power products in China to support inference for ultra-large-parameter models.

To date, the Zhenwu series of chips has served more than 650 enterprise customers across sectors including autonomous driving, finance, large models, embodied intelligence, energy, and manufacturing.

Alibaba said that as T-Head's chip product line matures and enterprise adoption expands, annual shipments of T-Head AI chips are expected to rise further.

T-Head Reveals Future Chip Roadmap

On the server CPU front, T-Head disclosed its roadmap for the Yitian series. The company plans to launch two generations of server CPUs, the Yitian 720 and Yitian 730, in 2027, with the Yitian 730 marking the first use of T-Head's fully in-house CPU microarchitecture.

The subsequent Yitian 750 will be based on a second-generation in-house CPU core and will support T-Head's proprietary ICN chip-to-chip interconnect bus protocol. This product can connect directly to the Zhenwu AI chips, further improving the coordination between CPU and AI chip.

T-Head also teased its next-generation AI chip, the Zhenwu J900. According to IT Home, the Zhenwu J900 will adopt a new in-house parallel computing architecture and is expected to launch in March 2028, though the exact timing remains subject to future official announcements.

Domestic AI Computing Power Enters an Era of Systems-Level Competition

As large-model parameter counts keep climbing, the focus of AI chip competition is shifting from peak per-chip computing power toward systems-level coordination across chips, interconnect, storage, servers, and software platforms.

With high-capacity memory, low-precision computing, and high-speed chip-to-chip interconnect, the Zhenwu V900 aims to address the computing power density, memory capacity, and communication efficiency challenges of large-model training and inference. The Panjiu super-node extends these chip capabilities to large-scale cluster deployment.

Looking at its product lineup, T-Head now spans core data center components including AI chips, server CPUs, smart network cards, interconnect chips, and SSD controllers. Going forward, competition among domestic AI computing power vendors will hinge not only on chip performance but also on super-node scale, software compatibility, cloud delivery capabilities, and real-world customer results.