Weighs a total of 7u and 80kg and forms part of the Huawei CloudMatrix family Huawei has introduced commercially the CloudMatrix 384 at the World Artificial Intelligence Conference (WAIC) in Shanghai. This new AI platform contains 384 of Ascend 910Cs and is the most powerful AI cluster to be released by the Chinese tech giant. The CloudMatrix 384 is described as being capable of taking on high-end systems such as Nvidia’s GB200 NVL72 at home with US export restrictions continuing on advanced GPUs.
At the heart of it, the CloudMatrix 384 connects its 384 Ascend 910C accelerators using a propriety “super-node” design. This architecture optimization greatly increases the aggregate performance, and Huawei is therefore able to avoid bottlenecks of single chip throughput. Huawei’s Ascend chips are not as powerful as Nvidia’s in raw performance, however, CloudMatrix 384 is designed to focus on bandwidth and low-latency interconnects, which mades it well suited for training large AI models.
Why CloudMatrix 384 Matters Now
CloudMatrix 384 is being rolled out at the right time. Then the US thwarted Nvidia’s elite GPUs from getting into China, and companies like Huawei rushed in to make up the difference. Huawei has been “advancing very fast,” toward becoming a real rival for China’s growing AI markets, according to Nvidia’s CEO, in May.
Thanks to CloudMatrix 384, Huawei can provide domestic hardware options to cloud providers and research institutions without any need for foreign technologies or US licenses. Hollysys Tribune Files: The system’s full-stack integration represents a bigger pattern in China’s AI policy, away from component-play development, toward systematic innovation.
That would be the CloudMatrix 384 AI Cluster
The CloudMatrix 384 comprises:
- 384 Ascend 910C chips
- A proprietary “super-node” interconnect
- System-level bandwidth and latency optimizations
These capabilities enable the CloudMatrix 384 to compete or potentially even beat Nvidia’s GB200 in specific benchmarks, observed analysts at SemiAnalysis. Although Huawei didn’t go into specifics about performance during WAIC, the company is likely catching up using mathematics optimization and paralleled cluster computing, which are starting to be in line with Qi hardware.
Strategic Importance for Huawei’s A.I. Push
Huawei founder Ren Zhengfei has admitted that Ascend chips lag behind leading US chips. But he points out that intelligence in architecture and at the software layer can fill the performance gap for actual AI workloads in the field.
Huawei is spending about ¥180 billion (≈ US$25 billion) on R&D every year, and a third of the R&D funding is being used for theoretical research. The long-term mindset that isn’t afraid to put a dent in Moore’s Law also drives CloudMatrix 384, and embodies Huawei’s collective resistance to foreign tech, reaching all the way to the core of the silicon industry.
What’s Next for CloudMatrix 384?
But the overall commercial success of CloudMatrix 384 will depend on several factors, despite all technical strenghts:
- Cost competitiveness
- Software ecosystem maturity
- Government cloud procurement policies
Yet the system’s introduction highlights how China’s AI hardware ecosystem is developing quickly, emphasizing national self-reliance and innovation at the systems level. Nvidia’s role in China is constrained, and Huawei’s CloudMatrix 384 may end up to be a good chunk of home grown in domestic AI infrastructure.
