China trials locally made AI computing system

China has begun beta testing its largest domestically produced AI computing node in Zhengzhou, Henan province.

The node uses three Sugon scaleX platforms with over 30,000 AI acceleration cards.

Its combined computing power is estimated at 15 exaflops based on platform specs.

Sugon is a publicly listed Chinese supercomputer firm.

The current world’s most powerful supercomputer, El Capitan, delivers 1.8 exaflops officially.

The node has an open architecture supporting multiple domestic accelerator brands and Nvidia’s CUDA ecosystem.

The trial shows China’s push for hyperscale AI infrastructure amid competition from firms like Baidu and Alibaba.

🧠 Food for thought

Implications, context, and why it matters.

SCNet and Sugon said the Zhengzhou node combines three Sugon scaleX platforms that can support deployments of more than 30,000 AI acceleration cards 1.
SCNet did not share the node’s total computing power. Using one scaleX as a baseline, three units could reach about 15 exaflops 1.
Sugon said parts of the scaleX supercluster already exceed capabilities claimed for Nvidia’s Rubin Ultra NVL576 rack, which Nvidia expects to introduce in 2027 2.
The effort puts Sugon up against domestic rivals such as Huawei, which has released the CloudMatrix 384 Supernode and plans products including the Atlas 950 SuperPoD 3.

Some descriptions of the Zhengzhou node mention an open architecture that can run multiple domestic accelerators plus Nvidia’s CUDA ecosystem, though the cited material does not spell out that open-architecture and CUDA support 1.
Separately, Sugon has teamed with more than 20 Chinese companies, including Lenovo and Moore Threads, to ship an open AI computing architecture pitched as an alternative to Nvidia’s CUDA 4.
Together, these moves aim to build a homegrown hardware and software stack that can keep AI work moving under US trade restrictions and cut long-term reliance on Western technology.