China builds domestic AI clusters
- China expanded an 'AI for Science' computing cluster in Zhengzhou using domestic accelerators and no US chips. (business-standard.com) - Nvidia's H200 chips reportedly remain unsold in China because Chinese firms face permission hurdles from Beijing. (reuters.com) - Those developments point toward parallel compute ecosystems being built for strategic and procurement reasons, not just export controls. (business-standard.com)
China has doubled the size of its Zhengzhou “AI for Science” computing cluster to about 60,000 domestically made accelerator chips. (business-standard.com) The Zhengzhou node began trial operations in early February with roughly 30,000 domestic accelerators and reached 60,000 by April 14, 2026, state media reported. (news.cgtn.com) Chinese supercomputer maker Sugon (Dawning Information Industry) supplied the accelerator cards and the system is sited at the national supercomputing internet’s core node in Zhengzhou. (scmp.com) Separately, U.S. officials say Nvidia’s advanced H200 AI chips have not yet been sold to Chinese firms; Commerce Department remarks on April 22, 2026, said Beijing has not approved purchases. (money.usnews.com) (reuters.com) U.S. reporting notes the Trump administration in January cleared some H200 sales with conditions, but shipments have been stalled by disagreements over licensing terms and Chinese approvals. (money.usnews.com) (reuters.com) Analysts and several reports characterize the two developments — fast domestic scaling in Zhengzhou and stalled foreign-chip sales — as evidence that parallel compute ecosystems are being built for strategic sourcing and procurement reasons, not only because of U.S. export controls. (business-standard.com) U.S. policy changed in January 2026 from a broad “presumption of denial” to case-by-case licensing for certain advanced AI chips, and Washington has also debated tightening an “affiliates” rule that would expand export limits. (money.usnews.com) (reuters.com) “AI for Science” (AI4S) refers to using AI models and specialized processors to speed complex simulations and discovery tasks — such as protein folding, materials design and climate modelling — that need large-scale compute. (ecns.cn) State outlets Xinhua and China Media Group flagged the Zhengzhou activation as a national infrastructure milestone, while U.S. officials have publicly warned about licensing and end‑use risks for advanced chips. (en.people.cn) The Zhengzhou cluster now runs on 60,000 domestic accelerators, and U.S. officials say H200s remain unsold in China as of April 22, 2026 — two supply‑chain trajectories likely to shape procurement and approvals in the months ahead. (news.cgtn.com)