May 1st News – According to foreign media reports, Huawei has successfully delivered advanced AI chip clusters, the CloudMatrix 384, to domestic customers affected by US export restrictions and unable to purchase Nvidia AI chip products, and has received an increasing number of orders from these customers. This chip cluster can efficiently connect a large number of chips (384 Ascend 910C) to form very powerful computing capabilities (which analysts say surpass Nvidia’s AI chip cluster NVL72).
According to information disclosed by two individuals familiar with the matter, Huawei’s CloudMatrix 384 chip has achieved remarkable sales performance, with over 10 units already sold. Among the first customers to receive this chip are data centers serving domestic technology companies, highlighting its market appeal.
Industry analysts have widely praised Huawei, stating that the speed at which it built and delivered the CloudMatrix 384 is astonishing. Dylan Patel, founder of the chip consulting firm SemiAnalysis, even stated directly that the emergence of Huawei’s CloudMatrix 384 signifies that China possesses an AI system capable of surpassing Nvidia’s. Information from Huawei’s company presentations shown to local clients and disclosures from knowledgeable sources indicate that the CloudMatrix 384 significantly outperforms Nvidia’s AI chip cluster NVL72 (composed of 72 Nvidia GB200 chips) in key performance indicators, with a computing capacity 67% higher and a total memory capacity more than three times greater. Patel explained that Huawei cleverly leverages advanced networking technology to compensate for the limitations of individual chip performance, significantly boosting the overall performance of the CloudMatrix 384 chip cluster.
However, the CloudMatrix 384 is not without its shortcomings. Compared to Nvidia’s similar products, it also has deficiencies. Due to the large number of chips used (384 Ascend 910C), the CloudMatrix 384 has higher energy consumption, leading to increased electricity costs. Furthermore, in terms of the software system, compared to Nvidia’s CUDA, Huawei’s system requires more effort from experienced engineers for maintenance, resulting in 3 to 5 times higher operational labor costs. On the other hand, it is worth mentioning that China has abundant power resources and a large pool of engineers. Under the circumstances of not being able to obtain Nvidia’s top-tier technology, the CloudMatrix 384 remains a highly viable option for domestic customers.
Foreign media reports indicate that a single CloudMatrix 384 unit is priced at approximately 60 million RMB (about 8.2 million USD), with the specific price subject to contract and potential fluctuations. Analysts estimate that Nvidia’s NVL72 is priced at around 3 million USD (Nvidia stated that its price range depends on OEM original equipment manufacturers and customer specifications). Despite the price difference, Huawei, with its continuous and strong technological innovation, has opened up new development paths for domestic enterprises in the AI chip field, breaking free from the shackles of external restrictions.