China's AI Firms Shift to 1,024-Chip Supernodes at WAIC as Systems Competition Deepens
Updated
Updated · jiemian.com · Jul 20
China's AI Firms Shift to 1,024-Chip Supernodes at WAIC as Systems Competition Deepens
3 articles · Updated · jiemian.com · Jul 20
Summary
Huawei’s Atlas 950 SuperPoD—shown physically in China for the first time at WAIC—packed 1,024 NPUs and claimed 1 exaFLOPS at FP8, underscoring the event’s turn from standalone chips to deployable AI systems.
That shift reflects rising inference demand and chipmaking constraints: vendors are trying to turn hundreds of accelerators into one usable machine, with customers focusing more on tokens per second and cost per token than peak theoretical performance.
ZTE, Alibaba Cloud, Lenovo and Sugon all highlighted supernode designs, but executives said interconnect bandwidth, latency, software adaptation and fault recovery now matter as much as GPU counts.
Most supernodes deployed in China still use fewer than 100 accelerator cards, and companies have yet to agree common interconnection standards, limiting how quickly larger systems can move from showpiece to broad adoption.
The broader contest is spreading across the supply chain, from chips and switches to fibre and cooling, as China tries to close the gap with Nvidia’s CUDA ecosystem through system-level optimisation.
As China builds a powerful, self-reliant AI ecosystem, what becomes the new global standard for technological innovation and leadership?
China's AI ambitions will consume vast energy. Can its renewable grid truly sustain this exponential demand without derailing climate goals?
China's ground data centers face overbuilding. Can its ambitious space computing network avoid becoming a multi-trillion dollar stranded asset?
From Sugon 8000 to Xingshu: How China’s 100,000-Card AI Supercluster and Satellite Network Are Redefining Global AI Leadership
Overview
At the 2026 World Artificial Intelligence Conference in Shanghai, China unveiled the Sugon 8000 (Dengfeng) AI Supercluster, immediately recognized as a 'Treasure of the Hall' for its groundbreaking scale and significance. This supercluster is fully domestically produced, highlighting China's commitment to technological independence in artificial intelligence. Its development reflects a major shift in China's chip industry, where innovators are adopting new approaches like software-defined design, logical folding, and 3D stacking. Together, these advances mark a pivotal moment for China's AI infrastructure, demonstrating both technical leadership and a drive for self-reliance.