[News] Alibaba T-Head Unveils Zhenwu M890 With 3× Performance; New AI Chips Planned for 3Q27/3Q28

As Chinese chipmakers accelerate efforts to develop domestic alternatives to NVIDIA, Alibaba’s T-Head has unveiled its latest AI chip roadmap. According to Mydrivers, at the 2026 Alibaba Cloud Summit, T-Head introduced the Zhenwu M890 AI chip for training and inference while also publicly disclosing the full roadmap for the Zhenwu chip series. Over the next two years, the company plans to launch the more powerful Zhenwu V900 and Zhenwu J900 to support AI computing demands in the emerging Agentic AI era.

As the report notes, the Zhenwu M890 integrates 144GB of HBM memory and delivers inter-chip interconnect bandwidth of up to 800GB/s. Its overall performance is said to be three times that of the previous-generation Zhenwu 810E. According to another Mydrivers report, the Zhenwu 810E, officially unveiled in January 2026, features a self-developed parallel computing architecture, 96GB of HBM2e memory, and inter-chip bandwidth of up to 700GB/s. Its overall performance was reportedly comparable to NVIDIA’s H20.

Looking ahead, Alibaba’s T-Head plans to launch the Zhenwu V900 in 3Q27. The chip is expected to feature a significantly upgraded in-house parallel computing architecture, deliver three times the performance of the Zhenwu M890, including 216GB of memory, and boost inter-chip bandwidth to 1,200GB/s. In 3Q28, Alibaba is set to introduce the Zhenwu J900, which is expected to bring a major leap in its self-developed parallel computing architecture and further enhance performance, the report adds.

Alibaba Unveils 128-Supernode AI System to Support Agentic AI

Meanwhile, at the summit, Alibaba also unveiled a 128-supernode server based on the Zhenwu M890, equipped with ICN Switch 1.0 and offering communication latency as low as the hundreds-of-nanoseconds level, Mydrivers notes. The system enables 128 AI chips to operate as a single computer, supporting massive concurrent agent inference and large-model training workloads.

Notably, T-Head also disclosed cumulative shipment figures for the Zhenwu chip series for the first time. As of the day of the summit, cumulative shipments had reached 560,000 units, serving more than 400 customers across over 20 industries, including China Telecom, FAW Group, and Shanghai Pudong Development Bank, the report notes.

In addition, Alibaba also launched its flagship large model, Qwen3.7-Max, according to Wallstreetcn. In a test, Qwen3.7-Max autonomously optimized a production-grade AI inference workload on T-Head’s Zhenwu M890 chip platform. After approximately 35 hours and 1,158 tool calls, the model achieved a geometric mean speedup of 10× compared with the reference implementation, the report notes.

Read more

• [News] Alibaba’s T-Head Reportedly Hits 470K Chip Shipments, Expands Amid Unclear IPO Timeline
• [News] Alibaba T-Head Unveils New AI Chip Said to Match NVIDIA H20 as IPO Speculation Builds

(Photo credit: Alibaba)