By Hamza Faisal
China's bid to out-compute the United States without American chips just got a showcase. The country's first fully domestically developed 100,000-chip AI supercluster has unveiled 10 major application achievements, proving — the builders say — that a system designed and manufactured entirely at home can do work previously exclusive to Nvidia's ecosystem.
Known as Sugon 8000, or Dengfeng (Peak), the machine was built by computer maker Sugon and announced in July at an intelligent computing conference in Zhengzhou. It has since been plugged into the national supercomputing internet, where it sells compute as a service rather than just hardware.
A hundred thousand chips in 2,000 square meters
The headline numbers are the ones that stop engineers mid-scroll. More than 100,000 AI chips and over 10,000 compute nodes packed into just 2,000 square meters, according to state broadcaster CGTN. Liquid cooling and high-speed interconnects keep the behemoth stable and efficient.
The cooling is the more remarkable engineering story. Sugon uses immersion phase-change liquid cooling, in which chips sit submerged in a coolant that boils at roughly 50 degrees Celsius. The approach cuts total energy consumption by about 40 percent and drives the cluster's power usage effectiveness (PUE) as low as 1.04 — world-class efficiency for a machine where a single AI chip can draw over 1,000 watts at full load.
The achievement, unveiled Sunday, is as much about independence as scale. The cluster's builders claim full-stack domestication: the chips, storage, networking and management software are all home-grown. Six core domestic chips — including processors from Hygon — form the foundation, while a self-developed distributed storage system took top spots on the 2026 IO500 benchmark and a native 400-gigabit lossless RDMA network, scaleFabric, compresses fault recovery from minutes to milliseconds.
Rockets, atoms and the universe
Among the 10 applications now public, the most striking is a partnership with Xi'an Jiaotong University that completed one of the largest fluid-dynamics simulations ever attempted. Using tens of thousands of domestic AI chips, researchers simulated the supersonic exhaust plume of a 33-engine reusable rocket array with a computational grid of 6.774 trillion cells — described as the largest known supersonic-jet simulation, and a tool for designing rockets that land themselves.
Other highlighted workloads span scientific computing and AI training on a unified infrastructure, from atomic-scale simulations to large language model inference. Sugon says the system already supports inference for more than 400 mainstream models, and its operators claim the cluster could serve as China's largest "token factory," covering 5 to 10 percent of the country's AI token demand.
The compute race sharpens
The unveiling lands days after a national report showed China's intelligent compute capacity reached 2,185 exaflops in the first half of 2026, up 177 percent year-on-year, with over 700 million generative-AI users. Microsoft, OpenAI and Meta are racing to build their own 100,000-card clusters; China's message is that export controls have not stopped it from joining the same club with its own supply chain.
The builders are already planning the sequel. Sugon has signed a strategic partnership with the Beijing Institute of Scientific Intelligence to develop a second fully domestic 100,000-card system — this time aimed at "super-intelligent convergence," merging supercomputing and AI into a single platform for scientific discovery.
By Hamza Faisal
