Huawei AI Chips: What CloudMatrix 384 Means for China
Huawei AI chips power CloudMatrix 384, a system linking 384 Ascend accelerators. See what its scale means - and what remains unproven.

Table of Contents
Quick Answer
Huawei CloudMatrix 384 links 384 Ascend 910C accelerators and 192 Kunpeng processors using Huawei’s UnifiedBus interconnect. It demonstrates Huawei’s effort to build large AI computing systems with domestic chips, but the reported configuration alone does not establish performance parity with competing platforms.
Why Huawei Is Linking Chips Into Larger AI Systems
CloudMatrix 384’s defining feature is its scale: instead of relying on a small number of the most powerful accelerators, it combines hundreds of Huawei chips into one computing system. That approach can pool processing capacity for demanding AI workloads, including model training and inference. It also gives Huawei a way to build larger systems around its own hardware rather than depending on unrestricted supplies of leading foreign chips.
Huawei’s strategy does not make the system a direct, like-for-like replacement for every competing platform. Performance depends on the workload, software and how effectively the processors work together. The reported configuration and interconnect identify how Huawei is assembling the system, but they do not establish that it matches other systems across all tasks.
What Adoption and AI Efficiency Can Tell Us
There are signs that Huawei hardware is being used for real AI development. Z.ai has said it trained models entirely on Huawei chips, according to WIRED. That is evidence of use, though one company’s experience does not show how widely the hardware has been adopted or how it performs across different workloads.
DeepSeek offers a related lesson about computing demand: more efficient training methods can help developers build competitive models with fewer resources. This matters because progress in AI does not depend only on obtaining the largest possible supply of advanced chips. Better use of available hardware can also help teams work within constraints, though efficiency gains do not eliminate the need for substantial computing capacity.
How Export Curbs Shape the Competition - and What Remains Unclear
US AI chip export controls were designed to limit Chinese access to certain advanced American chips. More recent policy efforts have also addressed how Chinese companies use chips, including hardware located outside China. These restrictions create an incentive for domestic alternatives such as Huawei’s Ascend systems, while complicating access to foreign hardware and the systems built around it.
The controls have not made Nvidia’s China business disappear. WIRED reported that China represented 8% of Nvidia’s revenue in the fiscal quarter that ended January 26, 2025. It also reported that sales in China of permitted older-generation chips and non-AI chips had doubled over the prior year. Those figures show continued business, but they do not by themselves measure how much advanced AI computing remains available to Chinese developers.
That overall impact is difficult to quantify. CloudMatrix 384 demonstrates Huawei’s push to assemble large domestic systems, while examples such as Z.ai’s training show some adoption. Neither establishes the full performance, reach or effect of export restrictions across China’s AI sector. Follow our coverage of AI chips and export controls as Huawei’s systems develop and China’s computing options change.
Key Statistics
- CloudMatrix 384 links 384 Ascend 910C accelerators.The article-reported system configuration describes Huawei’s approach to pooling domestic accelerators for large-scale AI computing.
- The reported configuration includes 192 Kunpeng processors.These processors are included alongside the 384 Ascend 910C accelerators in CloudMatrix 384.
- China represented 8% of Nvidia’s revenue in the fiscal quarter that ended January 26, 2025.WIRED reported this figure. It indicates continued Nvidia business in China but does not measure the availability of advanced AI computing there.
- Sales in China of permitted older-generation chips and non-AI chips had doubled over the prior year.WIRED reported this increase alongside Nvidia’s China revenue figure; it concerns permitted older-generation and non-AI products, not advanced AI chip access overall.
Frequently Asked Questions
What is Huawei CloudMatrix 384?
How many chips does Huawei CloudMatrix 384 link?
Why is Huawei building larger AI chip systems?
Has Z.ai trained models on Huawei AI chips?
Do Huawei AI chips match competing systems?
Key Takeaways
- CloudMatrix 384 combines 384 Ascend 910C accelerators with 192 Kunpeng processors.
- Huawei’s UnifiedBus interconnect links the system’s components.
- Large-scale chip pooling is one way Huawei aims to support demanding AI workloads with domestic hardware.
- Z.ai has said it trained models entirely on Huawei chips, indicating use but not broad adoption or performance across workloads.
- CloudMatrix 384 shows Huawei’s scaling strategy, but available details do not establish how it compares with other systems across tasks.