China brings a domestic 100,000-card AI cluster online

At Zhengzhou’s core node of China’s National Supercomputing Internet, the lights are on across China’s first fully domestic 100,000-card AI super-cluster. The system has entered service and is already supporting more than 300 workloads across twenty-six domains, from new materials to drug discovery.

This is not a single giant campus. A parallel cluster at the China Telecom Greater Bay Area hub is being used for trillion-parameter model training. Together, the two deployments are intended to form a national compute backbone for scientific workloads and large-scale AI training. Pandaily reports that train-to-completion cycles for a frontier-class model have been compressed from about a year to roughly six months.

The connective tissue is still being built. Engineers at Pengcheng National Laboratory are prototyping a hair-thin optical fiber with four internal channels, a 4x bandwidth upgrade over current single-mode fiber, for an intended first deployment on the Shenzhen-Guiyang line. Direct optical paths and latency targets are designed to make remote compute feel local to the user calling on it.

The software layer has begun trial operation, but the national fabric is not yet complete. More than 60 percent of China’s compute capacity is under unified monitoring, while eight national hub nodes are being folded into the East Data, West Compute framework. By the end of June, intelligent compute capacity had reached 218.5 EFLOPS, up 177 percent year-on-year.

So what changes in practice? Chinese AI labs and research teams can draw on a much larger pool of domestic compute for model training, industrial manufacturing and scientific work, with a second hub aimed at the largest models. The immediate limit shifts to coordination: whether the scheduling fabric can keep pace with the hardware already switched on. The fiber remains a prototype, and the monitoring platform is still in trial operation.