China's AI Supercluster: ScaleX "Ten Thousand Card" System Redefines AI Computing Power

Key Points

The Sugon (Zhongke Shuguang 中科曙光) scaleX “ten thousand card” supercluster is China’s first domestically produced “ten thousand card” AI cluster system, unveiled at HAIC2025.
It features the world’s first single-cabinet 640-card supernode, deploying 10,240 AI accelerator cards with a total computing power exceeding 5 EFlops.
The scaleX640 supernode uses immersion phase-change liquid cooling, boosting single-cabinet computing density by 20 times and achieving an ultra-low PUE of 1.04.
The system incorporates an independently developed native RDMA high-speed network (scaleFabric) with 400 Gb/s bandwidth and sub-1-microsecond latency, allowing for expansion to over 100,000 cards.
It boasts an intelligent operation and maintenance platform with 99.99% cluster availability and can increase AI accelerator card resource utilization to 55%.

Welcome to the future of artificial intelligence. China just unveiled a game-changer.

At the Guanghe Organization 2025 Artificial Intelligence Innovation Conference (HAIC2025) in Kunshan, a new titan emerged: Sugon’s (Zhongke Shuguang 中科曙光) scaleX “ten thousand card” supercluster.
This isn’t just another server rack; it’s a monumental leap in AI infrastructure, marking the first physical appearance of a domestically produced “ten thousand card” AI cluster system.
This supercluster promises to reshape how we think about computational scale and efficiency in the AI world.

What makes it so impressive? Its crown jewel is the world’s first single-cabinet 640-card supernode.
This design pushes single-cabinet computing power density up by an astonishing 20 times, easily outperforming current industry benchmarks.

Unpacking the ScaleX: China’s AI Powerhouse

Li Bin, Senior Vice President at Sugon (Zhongke Shuguang 中科曙光), highlighted the scaleX “ten thousand card” supercluster as a critical solution for tackling complex AI tasks.
Think trillion-parameter large models and advanced scientific AI (AI4S) – the kind of computing that pushes the boundaries of artificial intelligence.
Sugon’s extensive experience in large computer system design is clearly evident here.

The scaleX “ten thousand card” supercluster isn’t just bigger; it’s smarter.
It boasts multiple innovative breakthroughs, including:

Supernode architecture
High-speed interconnect networks
Storage performance optimization
System management and scheduling

These advancements directly address the extreme demands for performance, efficiency, reliability, and scalability in today’s AI infrastructure.
In fact, some of these technologies have already surpassed the 2027 NVL576 milestone outlined in the R&D roadmaps of similar overseas products.
That’s a significant head start.

1. The Groundbreaking Single-Cabinet 640-Card Supernode: A Game Changer for AI Computing Power

The heart of the scaleX supercluster is its unique design.
It comprises 16 Sugon (Zhongke Shuguang 中科曙光) scaleX640 supernodes.
These are linked together by the scaleFabric high-speed network, allowing for the deployment of a mind-boggling 10,240 AI accelerator cards.
The result? A total computing power exceeding an incredible 5 EFlops.

Let’s break down the scaleX640 itself:

It’s the world’s first single-cabinet 640-card supernode.
It uses ultra-high-density blade and immersion phase-change liquid cooling technologies.
This innovative cooling system boosts single-cabinet computing power density by 20 times.
It also achieves a remarkably low Power Usage Effectiveness (PUE) value of 1.04, which means almost all power consumed goes directly to computing, minimizing energy waste.
For context, a PUE of 1.0 means perfect efficiency, so 1.04 is exceptionally good.

Sugon’s investor relations activity report sheds more light on the scaleX640’s “one-to-two” high-density architecture.
This design enables ultra-high-speed bus interconnection for 640 cards within a single cabinet, creating a large-scale, high-bandwidth, low-latency supernode communication domain.
Think of it as an incredibly fast internal highway for data.

Furthermore, two scaleX640 supernodes can form a thousand-card computing unit.
Compared to competing products, the scaleX640 delivers a multiplicative increase in overall computing performance and that 20x boost in single-cabinet computing density.
This translates to significant real-world improvements, with a 30%-40% performance gain in MoE (Mixture of Experts) trillion-parameter large model training and inference scenarios, compared to traditional solutions.

And if you’re worried about reliability, don’t be. The scaleX640 has undergone over 30 days of stable operation reliability testing, ensuring it’s ready for ultra-large-scale cluster deployments.

The system is also built with an open architecture for AI computing.
This means hardware-wise, it supports multiple brands of accelerator cards, and software-wise, it’s compatible with mainstream computing ecosystems.
This flexibility establishes a new paradigm for domestic intelligent computing: “software-hardware collaborative and ecosystem-compatible.”
It’s primed to support cutting-edge scenarios like large model training, high-throughput inference, and scientific AI (AI4S).

2. Independently Developed Native RDMA High-Speed Network: Accelerating Data

Speed is paramount in AI, and the network is the backbone.
The Sugon (Zhongke Shuguang 中科曙光) scaleFabric network is built on China’s first 400G InfiniBand-like native Remote Direct Memory Access (RDMA) network card and switch chip.

This advanced network achieves:

Ultra-high bandwidth of 400 Gb/s
An incredibly low end-to-end communication latency of less than 1 microsecond.

This combination pushes communication performance between supernodes to an industry-leading level.
It fully unleashes the computing power of these “ten thousand card” superclusters and,
crucially, allows for the expansion of supercluster scale to over 100,000 cards.
Imagine the possibilities with that level of scalability!

3. Deep Optimization of Storage, Compute, and Transmission Coupling: Data Flow Excellence

It’s not enough to have powerful cards; the data needs to flow seamlessly.
The scaleX “ten thousand card” supercluster achieves this with designs like “super tunnel” and AI data acceleration.
This provides three-level data transmission synergy optimization:

Chip-level
System-level
Application-level

This ensures the storage platform can efficiently meet the strict bandwidth demands of “ten thousand card” concurrent read/write operations during extensive large model training.
What does this mean for users?

Improved response speed
Enhanced result accuracy during high-throughput AI inference
AI accelerator card resource utilization can be increased to 55%, meaning less wasted compute power.

4. Supercluster Digital Twin and Intelligent Scheduling: Smarter Management

Managing a cluster of this magnitude requires advanced tools.
The scaleX system employs digital twins of physical clusters.
This allows for visualized intelligent management, streamlining the entire process of cluster fault localization and repair.

Key features include:

An intelligent operation and maintenance platform that supports long-term cluster availability of up to 99.99%.
That’s less than an hour of downtime per year!
An intelligent scheduling engine capable of managing tens of thousands of nodes.
It can serve hundreds of thousands of users.
It supports tens of thousands of job schedules per second, ensuring that this massive computing power is utilized efficiently and effectively.

Find Top Talent on China's Leading Networks

Post Across China's Job Sites from $299 / role
Qualified Applicant Bundles
One Central Candidate Hub

Get 20% Off
Your First Job Post Use Checkout Code 'Fresh20'

Sugon’s Vision: Future Strategic Planning

Sugon (Zhongke Shuguang 中科曙光) sees its future driven by three core pillars:

National strategic dividends
The construction of technological barriers
Surging market demand

The company is strategically positioned for significant growth in key areas:

Domestic computing power
AI computing power expansion
Breakthroughs in computing technology architecture

Their global expansion and collaborative ecosystem initiatives are also broadening their long-term growth trajectory.
Sugon plans to stay focused on its core business of high-end computers, continuously innovating with cutting-edge technologies like supernode intelligent computing power, scientific large model development platforms, and supercluster systems.

The goal is to expand their full-stack capabilities in intelligent computing, computing power scheduling, and data center solutions.
This aims to perfect the entire industry chain system, from “chip to device, cloud to computation.”

The Impact of China’s “Ten Thousand Card” Supercluster

This debut isn’t just a technical achievement; it’s a strategic move in the global AI race.
By developing such a powerful, efficient, and scalable AI infrastructure domestically, China is bolstering its technological independence and leadership in AI development.
For investors, founders, and tech enthusiasts, this indicates a strong commitment to advancing large-scale AI capabilities.
Expect to see this technology underpinning rapid advancements in everything from LLMs to complex scientific simulations.

The scaleX “ten thousand card” supercluster is more than just hardware; it’s a testament to innovation, scale, and the relentless pursuit of computing excellence.
This is what pushing the boundaries of AI looks like.