High-Performance Computing via the Cloud

Alibaba Cloud has officially initiated commercial sales for its Lingjun Zhenwu M890 supernode, providing a breakthrough for organizations looking to leverage advanced AI power without the need to construct their own physical data centers. This new offering allows clients to rent a 64-card, high-speed computing cluster on an hourly basis.

This development represents a significant milestone, as it is the first time a Chinese-developed supernode has been made available directly through a public cloud platform. Unlike traditional instances that require complex pre-configuration or long-term purchase agreements, this service aims to provide a streamlined, enterprise-centric solution.


Technical Capabilities and Architecture

A supernode functions by linking multiple accelerators so tightly that software perceives them as a single, massive processor rather than a collection of separate chips. This architecture is vital for training trillion-parameter models, as it prevents the significant latency overhead that typically occurs when data travels between chips over slower networks.

The Zhenwu M890 is powered by technology from T-Head, Alibaba's internal chip design unit. Key specifications include:

  • Memory: 144GB of HBM memory per chip.
  • Bandwidth: Inter-chip bandwidth reaches 800 GB/s.
  • Scalability: Utilizing the proprietary ICNSwitch 1.0, the system expands the scale-up domain from the previous generation's 16 cards to a robust 64-card configuration.

According to the company, this setup is theoretically capable of managing a Mixture of Experts (MoE) AI configuration of up to 10 trillion parameters and is already being used to support commercial services like Kimi K3 and Qwen3.8-Max.


Strategic Context and Future Outlook

The emergence of the M890 reflects a broader shift toward technological self-reliance within China. As international restrictions limit access to certain high-end foreign processors, Alibaba’s use of homegrown silicon serves as a critical alternative for domestic AI startups and enterprise clients. While competitors like Nvidia’s GB200 NVL72 and AMD’s Helios configuration remain global leaders, Alibaba’s offering provides a distinct advantage to Chinese customers through lower latency and a localized infrastructure.

Alibaba has already outlined an ambitious roadmap for future hardware:

«The company expects the Zhenwu V900 to arrive in Q3 2027, promising triple the performance of the M890, 216GB of memory, and 1,200 GB/s of bandwidth, with the J900 series to follow in 2028.»

Currently, the M890 is restricted to users within the Inner Mongolia Autonomous Region. While pricing and energy efficiency metrics remain undisclosed, the deployment highlights China's growing capabilities in the AI hardware sector and its intent to expand such offerings to a broader market over time.