Alibaba’s Zhenwu M890 and Qwen3.7‑Max Signal China’s Push for a Domestic AI Hardware Stack

The chip is specifically designed to support the emerging generation of AI agents, which require high memory capacity and fast communication between processors to coordinate complex tasks and maintain long context windows.

High-Speed Interconnect for Large AI Systems

Beyond raw compute, Alibaba emphasized system‑level scaling capabilities in the M890 architecture.

The processor uses a self-developed parallel computing architecture with up to 800GB/s chip‑to‑chip interconnect bandwidth, enabling fast data exchange between accelerators in multi‑GPU clusters.

Alibaba also introduced ICN Switch 1.0, a networking chip that enables 64-card full‑bandwidth interconnection, a critical capability for distributed training of large models and for handling high‑concurrency inference requests from AI agents.

This architecture is designed to reduce communication bottlenecks that often limit performance in large AI clusters.

The Panjiu AL128 Hyperscale Server

To turn its chip into deployable infrastructure, Alibaba unveiled the Panjiu AL128 hyperscale server node, built around the Zhenwu M890.

The system integrates 128 AI accelerators into a single node, using ICN Switch 1.0 to connect them with extremely low latency. This configuration reduces the communication overhead typically associated with distributed computing clusters.

Such dense AI nodes are intended for:

• Large-scale model training
• High‑concurrency inference
• AI agent orchestration workloads

By designing both the chips and the server architecture, Alibaba can optimize performance across the entire stack rather than relying on third‑party hardware.

Qwen3.7‑Max: A Model Built for AI Agents

Alongside the hardware launch, Alibaba introduced Qwen3.7‑Max, the latest version of its Qwen large language model.

The model is positioned as a foundation model for agentic AI, with strengths in:

• Agentic coding and tool use
• Complex reasoning tasks
• Long‑horizon workflows, where the system must plan and act across extended sequences of steps.

According to Alibaba, Qwen3.7‑Max can sustain long-running autonomous operations—reportedly performing tasks for up to 35 hours without performance degradation.

This type of capability is important for AI agents that perform extended workflows such as software development, data analysis, or operational automation.

A Full-Stack AI Strategy

The announcements collectively illustrate Alibaba’s attempt to build a full-stack AI platform, combining several layers:

• Custom AI chips (Zhenwu series)
• High‑density server infrastructure (Panjiu systems)
• Cloud AI services on Alibaba Cloud
• Foundation models like the Qwen family

This mirrors the vertically integrated approach used by Nvidia, where hardware, networking, and software are tightly integrated to maximize performance for AI workloads.

Alibaba has framed the architecture around the “agentic era” of AI—where models no longer simply answer questions but instead act autonomously, plan multi-step tasks, and coordinate with external tools and software systems.

The Roadmap: V900 and J900 Chips

The company also revealed a forward-looking chip roadmap for the Zhenwu series:

• Zhenwu V900, planned for release in 2027
• Zhenwu J900, expected in 2028

The roadmap signals a long-term effort to continually improve domestic AI silicon performance and compete with global accelerator ecosystems.

Why This Matters for China’s AI Industry

Alibaba’s announcements highlight a broader strategic shift across China’s technology sector. With U.S. export controls limiting Chinese access to advanced Nvidia processors, companies are accelerating efforts to develop homegrown AI chips and infrastructure.

If successful, Alibaba’s approach could provide Chinese developers with a complete AI platform—from chips to cloud services—capable of supporting large‑scale training, inference, and autonomous AI agents.

However, many of the performance claims around the new hardware currently come from company announcements or media reports based on those disclosures. Independent benchmarks comparing the Zhenwu M890 with Nvidia accelerators such as the H100 or newer chips have not yet been widely published.

Even so, the launch signals a clear direction: major cloud providers are racing to control every layer of the AI stack, and Alibaba is positioning itself as a central player in China’s domestic AI infrastructure ecosystem.