Huawei has officially unveiled a massive AI infrastructure system, the CloudMatrix 384 Supernode, which it claims rivals and outperforms Nvidia’s flagship NVL72 GPU cluster in key areas.
Reported by the South China Morning Post and confirmed by SemiAnalysis, the new system marks a notable step in Huawei’s bid to become China’s answer to Nvidia amid ongoing US export controls.
The CloudMatrix 384 system connects 384 Ascend 910C chips, producing up to 300 petaflops of dense BF16 compute. In contrast, Nvidia’s NVL72 system, which links 72 GPUs via its proprietary NVLink technology, reaches around 180 petaflops. Huawei insiders have even dubbed it a “nuclear-level product,” signaling the strategic importance of this launch for China’s AI future.
Huawei’s new supernode boasts 3.6 times more aggregate memory and 2.1 times more memory bandwidth than Nvidia’s NVL72. That said, it draws nearly four times the power, a metric that would raise red flags in Western data centers—but less so in China. Thanks to China’s abundant and expanding energy grid, Huawei can afford to prioritize brute-force performance over efficiency.
The first CloudMatrix 384 racks are already live in Huawei’s data centers in Wuhu, Anhui province. The company’s engineering strength—especially in networking, optics, and system integration—has helped it sidestep export restrictions on chips and components by leveraging complex global sourcing strategies.
While the Ascend 910C chips still depend on foreign suppliers like Samsung (HBM) and TSMC (wafers), Huawei’s success in launching a system that outmuscles Nvidia in compute and memory shows how far China has come in developing self-reliant AI infrastructure. The CloudMatrix 384 may not be as energy-efficient, but for Beijing, the trade-off may be worth it, especially as AI becomes central to national competitiveness.