Fully domestic 100,000-card AI supercluster goes live in China
China's first fully domestic 100,000-card AI supercluster is now live. The National Development and Reform Commission confirmed that the Sugon-built Dawning 8000 is in operation at the National Supercomputing Internet's Zhengzhou core node, running 100,000 domestically produced AI accelerator cards across the full precision range from scientific to AI computing. Its peak throughput is equivalent to all of humanity computing nonstop for 200 years, and it already serves 300-plus tasks across 26 fields — new materials, drug discovery, and AI among them — according to Bu Jingde, chief engineer at the National High-Performance Computer Engineering Technology Research Center.
The milestone moves Chinese model training from 10,000-card experiments to 100,000-card engineering. NDRC data shows national intelligent computing capacity grew 177% year over year in the first half, and domestic reports say Sugon has already started building a second fully domestic 100,000-card system. China Telecom's Greater Bay Area node, meanwhile, has turned on a new domestic 10,000-card cluster that its builders say can cut trillion-parameter training time from a year to about six months.
This is the supply-side answer to US export controls — proof that China can scale frontier-class training without foreign accelerators. The question now is utilization, not buildout: the cluster reportedly ran at full capacity in its first week, but analysts already flag inland compute parks running at under 50% capacity, and the national dispatch network meant to balance that oversupply is still being stitched together.
QuantHealth raised $45 million in Series B funding to keep simulating clinical trials before they run. The Tel Aviv startup, which models trial outcomes in silico before a single patient enrolls, took the round led by Qumra Capital with participation from Sanofi Ventures, Pitango HealthTech, Bertelsmann Healthcare Investments, and others, bringing total funding to roughly $70 million. QuantHealth says it has simulated 600-plus trials across 30 indications with up to 90% predictive accuracy, and that 12 of the world's top 20 pharma companies already use its software; the new money will expand coverage to more than 40 indications and grow the team. It's a bet on the least-automated corner of drug development — about 90% of drugs entering trials fail, and most of that money is burned before anyone learns why. Simulation won't fix bad science, but it attacks the industry's most expensive feedback loop.
What to watch: whether Zhengzhou's cluster stays near full load as the national compute grid comes online — and whether China's 100,000-card buildout becomes the template for every hub node.
China's compute buildout is racing ahead of demand in places — efficiency win or bubble? Tell us in the comments.
