Huawei has begun delivering its advanced AI chip "cluster" to Chinese customers who are increasing orders after being cut off from NVIDIA's semiconductors due to export restrictions in Washington.
According to two knowledge about the matter, the Shenzhen-based technology group has sold 10 sets of CloudMatrix 384s, connecting a large number of chips together.
Those who received the first delivery included data centers that served Chinese tech companies, one of the people said.
Industry analysts said they were impressed by the speed at which Huawei built and began shipping CloudMatrix, which is connected to 384 AI processors to provide the computing power needed to develop AI models and services.
"The development of Huawei's CloudMatrix 384 means China now has an AI system that can beat Nvidia's," said Dylan Patel, founder of Chip Consultancy semi-analytics.
Huawei's new product is because Chinese customers are further restricted by NVIDIA's AI chips. The U.S. panel said that President Donald Trump will need $5.5 billion in revenue this month after making his H20 chip, tailored to comply with earlier export controls and requires special licenses to be sold to Chinese customers.
Huawei is accelerating its chip development to help domestic tech companies compete with their global peers, even though they limit opportunities for the best technology in the world in the region.
It has told local customers that according to the performance of cloudmatrix of the US tech giant (NVL72), NVIDIA's NVL72 is a popular AI cluster used by the US tech giant, which consists of 72 GB200 chips in terms of computing power and memory.
The CloudMatrix 384 uses Huawei's Ascend 910c chip, which underperforms NVIDIA's GB200 processor.
But Huawei uses a large number of chips (through its "supernode" connections, all processors connected using optical technology can be connected via optical technology - to improve the overall performance of the cluster.
In the company's speech, Huawei claimed that CloudMatrix performed better than NVL72 in key metrics, saying its calculations were 67% higher and its total memory capacity was more than three times larger.
"This can compensate for weaker personal chip performance through advanced networks to improve chip performance in clusters," Patel said on semi-analysis.
Industry analysts say Huawei uses its background on telecommunications to improve the performance of its chip systems.
Industry experts say that despite strong metrics, CloudMatrix 384 has several drawbacks compared to NVIDIA's competitors.
Due to the use of more chips, the CloudMatrix 384 has higher energy consumption, resulting in greater electricity bills. Compared with NVIDIA's CUDA, Huawei's software systems require more maintenance from experienced engineers, resulting in three to five times the cost of operating labor.
Nevertheless, given China's rich power and engineers, CloudMatrix is a viable solution for customers, especially when they are isolated from NVIDIA's state-of-the-art technology.
According to industry insiders, the CloudMatrix 384 will be around RMB 60 million (US$8.2 million), and the exact price depends on the individual contract. According to analyst estimates, this is compared to about $3 million for NVIDIA's NVL72.
NVIDIA says it has a high price range based on OEM and customer specifications.
Huawei declined to comment.