Alibaba Tuhao M890 Debuts with Triple Performance, Ushering in Full-Stack Agent Era for Chip-Cloud-Model-Inference

On May 20, 2026, at the Alibaba Cloud Summit, Alibaba Cloud announced the completion of a full-stack technology system upgrade designed for the Agentic era. The transformation reshaped the entire pipeline—from underlying chips and cloud platform to models and inference solutions. This shift positions Alibaba Cloud as an AI factory that enables 24/7 continuous operation of massive agents, moving beyond serving human users directly.

1. Core Foundation: Tengxun Zhenwu M890 Chip and Super Node Server

At the core of this upgrade is Tengxun's next-generation AI chip, the Zhenwu M890, which integrates training and inference.

Performance Improvement: The M890 features 144 GB of memory and delivers three times the performance of its predecessor, the Zhenwu 810E. It natively supports a range of data precision levels from FP32 to FP4, making it ideal for high-precision training and ultra-low-precision concurrent inference in agent scenarios.

Cluster Interconnection Breakthrough: Through integration with the proprietary ICN Switch 1.0 interconnect chip, Alibaba Cloud released the Panjiu AL128 Super Node Server, built on the Zhenwu M890. This server coordinates storage, computing, and networking across 128 AI chips at the system level, achieving nanosecond-level communication latency and substantially boosting the efficiency and stability of large-scale intelligent computing clusters.

Future Plan: Tengxun publicly disclosed the Zhenwu series chip roadmap for the first time, confirming that the Zhenwu V900 and Zhenwu J900 will be launched within the next two years, reinforcing its long-term competitiveness in the data center computing market.

2. Core Access Point: Reimagined "Qwen Cloud" and Agent-Centric Interaction

Alibaba Cloud fundamentally transformed cloud interaction logic. While traditional cloud platforms were built for humans—with control panels and dashboards—the cloud of the Agentic era must be designed for agents.

AI-Native Website "Qwen Cloud": The AI-native website Qwen Cloud replaces the traditional complex product catalog with a standardized Skills installation code. Agents can directly parse these code instructions and autonomously invoke computing, storage, and model capabilities, eliminating the need for manual control panel configuration.

Standardization of Capabilities: Alibaba Cloud has packaged over 150 mainstream models and cloud product capabilities into standardized Skills and CLI tools. With a single line of instruction, tools like Claude Code and mainstream agent frameworks can quickly install and access the full range of Alibaba Cloud's infrastructure capabilities.

3. Technical Strategy: Full-Stack Integration of Chip, Cloud, Model, and Inference

This new system is designed to handle the unique challenges of agent workloads, characterized by irregular elasticity, short life cycles, and extremely high instantaneous concurrency.

Deep Optimization: Alibaba Cloud not only offers models like the flagship Qwen3.7-Max but also achieves optimal computing resource scheduling through deep integration between the underlying Zhenwu series chips and the inference framework.

Shift in Objectives: As Alibaba Cloud CTO Feifei Li and other experts noted, the focus of large models has shifted from aligning with human preferences—saying things well—to aligning with task objectives—getting things done. The entire system evolution ensures that agents can efficiently complete complex engineering tasks in milliseconds, lowering the barrier to AI adoption across industries.

Summary:

By combining its Tengxun chip matrix with the Qwen Cloud access point and full-stack model inference, Alibaba Cloud has become the first in the industry to transition from a computing power rental provider to an AI factory. This system not only provides infrastructure to support the explosive growth of agents but also showcases the ambition of Chinese tech leaders to reshape the global productivity gateway through hardware-software collaboration in the Agentic era.