Catenaa, Friday, May 02, 2025-Chinese tech giant Huawei has begun deliveries of its most advanced artificial intelligence chip cluster, the CloudMatrix 384, to local clients cut off from US semiconductor leader Nvidia due to Washington’s tightening export restrictions, according to sources familiar with the development.
The high-performance system, which links 384 AI processors using Huawei’s proprietary “super node” optical networking, has already been delivered to more than 10 customers, including major data centers servicing Chinese tech firms. Analysts say Huawei’s rapid development and deployment of the cluster marks a significant stride in China’s effort to close the AI chip gap with the US.
The CloudMatrix 384, built with Huawei’s Ascend 910C chips, is touted to outperform Nvidia’s NVL72 in total computing power and memory capacity, according to internal company presentations reviewed by the *Financial Times*. Though each individual Huawei chip lags behind Nvidia’s GB200 processor, the system’s scale and optical interconnect compensate for the performance gap.
Huawei claims the system is 67% stronger in computing power and offers triple the memory capacity of NVL72. Still, the cluster comes with tradeoffs: greater energy use and higher operational costs due to its complex software environment and hardware footprint.
With Nvidia’s H20 chip now requiring a special export license, Huawei’s CloudMatrix — priced around $8.2 million per set — has emerged as a domestic alternative, albeit one that demands significant power and engineering resources.
