en.Wedoany.com Reported - Alibaba's semiconductor company T-Head officially launched its new generation training-inference integrated AI chip, the Zhenwu M890, at the 2026 Alibaba Cloud Summit held on May 20. The chip features 144GB of HBM memory, an inter-chip interconnect bandwidth of 800GB/s, and overall performance three times that of its predecessor, the Zhenwu 810E. It natively supports multiple data precisions from FP32 to FP4, covering full-scenario needs from high-precision training to ultra-low precision inference. The Zhenwu M890, combined with the self-developed ICN Switch 1.0 interconnect chip, enables full-bandwidth interconnection of 64 cards, significantly improving the computational efficiency and stability of large-scale intelligent computing clusters.
The performance leap of the Zhenwu M890 is built upon a clear generational comparison with its predecessor. The Zhenwu 810E was launched in the second quarter of 2024, equipped with 96GB of HBM2e memory and an inter-chip interconnect bandwidth of 700GB/s. The Zhenwu M890 increases memory capacity by 50%, boosts interconnect bandwidth by approximately 14%, and achieves a threefold leap in overall card performance through architectural upgrades. Alibaba Cloud simultaneously released the Panjiu AL128 supernode server based on the Zhenwu M890. Equipped with the ICN Switch 1.0, it allows 128 AI chips to form a single computer with P2P communication latency under 150ns, primarily designed to address massive concurrent inference and large model training needs in Agent scenarios. This server is now available on the Alibaba Cloud Bailian Platform, supporting mainstream models like Qwen, DeepSeek, and Kimi.
The release of the Zhenwu M890 is not an isolated event but a core part of Alibaba Cloud's reconstruction of its full-stack technology system for the Agentic era. On the day of the summit, Alibaba Cloud announced the completion of its "Chip-Cloud-Model-Inference" full-stack Agentification upgrade, simultaneously launching "Qianwen Cloud," a new official website for AI products born for Agents, and the latest flagship large model, Qwen3.7-Max. Liu Weiguang, Senior Vice President of Alibaba Cloud, stated that once Agents break through the critical point, they can work 24/7, creating endless demand for AI and cloud services. Alibaba Cloud is undergoing full-stack technological innovation, upgrading from underlying chips, Agentic Cloud, models to inference platforms, to build China's largest AI factory.
At this summit, T-Head publicly revealed the complete product roadmap for the Zhenwu series chips for the first time. Gao Hui, Vice President of T-Head Semiconductor, disclosed that the release cadence for future Zhenwu series chips will accelerate to a "one generation per year" rhythm: the Zhenwu V900 will be launched in the third quarter of 2027, featuring a deeply iterated self-developed parallel computing architecture with performance three times that of the Zhenwu M890, equipped with 216GB of memory and an inter-chip interconnect bandwidth increased to 1200GB/s; the Zhenwu J900 is scheduled for release in the third quarter of 2028, achieving a leapfrog innovation in the self-developed parallel computing architecture.
Gao Hui revealed at the summit that as of April 2026, cumulative shipments of the Zhenwu series chips had reached 560,000 units, serving over 400 customers across more than 20 industries, including China Telecom, FAW Group, and SPD Bank. Compared to the "cumulative large-scale delivery of 470,000 units as of February 2026" disclosed by Alibaba during its March earnings call, shipments increased by 90,000 units in two months. By industry, Zhenwu AI chips have been deployed with over 130,000 cards in the intelligent driving sector, serving more than 30 leading customers including Changan Automobile, GAC Group, BYD, XPeng, NIO, and Li Auto, with verified compatibility with over 50 mainstream autonomous driving models; in the financial industry, 100,000 cards have been deployed, serving over 150 customers. A report released by market research firm IDC in April shows that in China's cloud AI accelerator market in 2025, ranked by shipment volume, T-Head ranked second among domestic AI chip manufacturers with 265,000 units.
The Qwen3.7-Max flagship model, also unveiled at this summit, demonstrated performance levels close to the strongest versions of GPT, Claude, and Gemini across multiple benchmarks. In reasoning capability, it scored 92.4 on GPQA Diamond; in coding agents, it scored 80.4 on SWE-Verified; and on the general agent benchmark MCP-Mark, it scored 60.8. The model can autonomously complete over 1,000 tool calls within 35 consecutive hours, possessing durable and stable long-cycle execution capabilities, making it one of the most representative long-horizon agent foundation models currently available.
The release of the Zhenwu M890 is highly consistent with Alibaba Group's overall AI strategic investment. Alibaba Group CEO Wu Yongming stated during the May 13 earnings call that the annualized recurring revenue (ARR) from AI models and application services has exceeded 8 billion RMB, and is expected to surpass 30 billion RMB by year-end. Last year, Alibaba committed to investing over 380 billion RMB (approximately 53 billion USD) in cloud and AI infrastructure over the next three years. The full-stack technology system released at this summit represents a phased realization of this strategic layout. T-Head has been spun off from Alibaba Group and is preparing for an independent IPO. As the iteration rhythm of the Zhenwu series chips compresses from two years per generation to one year per generation, T-Head is accelerating the expansion of its self-developed chips' penetration in the AI computing market, providing underlying computing support for Alibaba Cloud's Agentic era strategy.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
