Published on Jan. 02, 2026

China’s first physically reveiled 10,000-card AI supercluster is reported to deliver more than 5 EFLOPS of total computing power, marking a breakthrough in the country’s push to build advanced hardware, reports Semi Insights (in Chinese).

At the launch event on December 17, 2025, Sugon said its cluster, called scaleX, is much more effienct in training and inference compared with conventional approaches. This would make it suitable for trillion-parameter AI models.

Sugon, officially Dawning Information Industry Company Ltd., is based in Beijing’s Zhongguancun district, sometimes referred to as China’s Silicon Valley. Founded by researchers affiliated with the Chinese Academy of Sciences, the company produces supercomputers and IT infrastructure. It is listed on the Shanghai Stock Exchange.

According to the company, the new system builds on earlier developments and features what it calls the world’s first 640-GPU supernode per cabinet. Sixteen such supernodes are interconnected, deploying 10,240 accelerator cards, delivering more than 5 EFLOPS of total computing power, and achieving a PUE as low as 1.04.

By connecting domestically produced GPUs into ever-larger clusters, Chinese firms are pursuing an alternative strategy to Western rivals, effectively working around U.S. export restrictions that restrict the sale of both the most advanced AI chips and the lithography machines needed for their production to China.

The cluster approach is also driven by Moore’s law approaching physical limits, necessitating new ways to design hardware that keeps pace with rapidly growing data-center demand. Single nodes are no longer capable of meeting the computing requirements of artificial intelligence.

"While the U.S. focuses on ever more powerful individual chips from firms such as Nvidia, China is taking a different approach, pursuing comparable results through large-scale system design and software optimization,” writes the Chinese business weekly Caixin. (In Chinese).

"With advanced foundry access restricted, Chinese firms are leaning on cluster-based architectures that interconnect thousands of lower-powered domestic chips with high-speed links, to achieve competitive system-level output, albeit at a higher energy cost", Caixin writes.

Huawei is leading this approach. After being placed under U.S. sanctions in 2020, its engineers developed the concept of “supernodes” integrating thousands of Ascend chips on unified platforms.

Huawei's Atlas 900 supernode, launched in March 2025, connects 384 Ascend 910C chips for a peak performance of 300 PFLOPS (floating-point operations per second). At the time, this was the world’s most powerful known AI compute node.

Huawei has already announced even more powerful future versions that will scale dramatically, with the Atlas 950 due to become available in 2026, linking 8,192 chips, and the Atlas 960, due in 2027, connecting over 15,000. And this does not have to be the end.

While each Huawei Ascend 910C chip offers roughly one-third of the performance of Nvidia’s latest GPUs, a comparison by Semi Analysis found that Huawei’s clustered approach can deliver nearly twice the system-level performance.

This modular cluster design theoretically allows for further expansion to hundreds of thousands of chips to train ultra-large AI models, although technical hurdles will have to be overcome while scaling the clusters.

At both Huawei and Sugon, developers acknowledge that the software ecosystem remains the biggest challenge in gaining market share from Nvidia. Their answer is to consider open source solutions.

"Our hope is to use an open architecture to make the technologies we have accumulated available across the entire value chain, so that our industry partners can focus on what they do best within this ecosystem, and together we can get this done,” Li Bin, Senior Vice President of Sugon, said in a speech during the launch of his company's 10,000-GPU AI supercluster.