Huawei's SuperCluster targets Nvidia alternatives as China seeks self-reliance

Huawei Technologies Co announced on Thursday the launch of what it claims are the world’s most powerful AI computing clusters: SuperNode and SuperCluster, despite ongoing US sanctions. This marks the company’s latest development in high-performance computing.

“Computing power has been and will continue to be central to artificial intelligence, particularly for China’s AI development,” said Xu Zhijun, Huawei’s rotating chairman, during a speech in Shanghai.

As China seeks to enhance its domestic AI capabilities and reduce reliance on western technology providers, Huawei is introducing high-performance processing architectures designed to allow processors to connect at extremely high speeds. The move reflects broader national efforts to build self-sufficiency in critical technologies.

Advanced semiconductor technologies remain a key point of tension in the ongoing geopolitical rift between China and the United States. In recent years, the US has imposed restrictions on the export of AI chips to Chinese firms, including Huawei, in a bid to limit their access to cutting-edge computing power.

In response, Huawei plans to release three new versions of its Ascend chip between now and the end of 2028. According to the company, each iteration aims to "double compute" performance, forming the backbone of its AI infrastructure.

Huawei’s AI computing architecture is built on superclusters composed of numerous superpods, which are themselves made up of multiple supernodes. These supernodes are built using Ascend chips and rely on system-level innovations to circumvent technical constraints imposed by sanctions.

The company stated that its new Atlas 950 SuperCluster will incorporate more than 500,000 processors, while each Atlas 950 SuperNode will support 8,192 Ascend chips. A SuperPod is described as a single logical unit formed from several physical machines capable of collective reasoning, learning, and decision-making.

Huawei introduced two new SuperPod models: the Atlas 950SuperPod, which includes 8,192 Ascend neural processing units (NPUs), and the more powerful Atlas 960 SuperPod, with 15,488 Ascend NPUs. These systems are designed to lead the industry across several dimensions, including the number of NPUs, overall processing capacity, memory bandwidth, and interconnect speed.

According to Huawei, these SuperPods represent the most powerful AI computing systems currently available globally, a claim it says is supported by publicly available roadmaps from industry competitors.

Looking ahead, Huawei plans to release a more advanced version of the Atlas 960 in 2027, also containing 15,488 Ascend chips per node. The full supercluster will eventually contain over one million Ascend chips, according to the company’s roadmap.

To date, Huawei has delivered more than 300 Atlas 900 A3 supernodes to over 20 customers across sectors such as manufacturing and telecommunications.

Huawei’s SuperPod architecture appears to be a strategic effort to compete with Nvidia’s NVLink technology, enabling high-speed communication between server CPUs. Given that Nvidia’s AI silicon currently outperforms Huawei’s most advanced Ascend chips, Huawei’s system architecture plays a key role in narrowing the performance gap.

The timing of the launch aligns with China’s broader push to develop domestic alternatives to Nvidia’s offerings. On Monday, Chinese regulators announced an expanded investigation into Nvidia for alleged monopolistic practices, adding to mounting pressure on foreign AI chip providers.

Huawei’s latest move reflects a growing trend among Chinese firms to develop self-reliant solutions in the face of US technology restrictions. While it may not represent a breakthrough in chip design, the approach underscores the country's strategic efforts to build a sustainable AI ecosystem using domestically developed infrastructure.