As AI models expand beyond 1 trillion parameters, a new concept has come to dominate China’s computing landscape: the supernode. At this year’s World Artificial Intelligence Conference (WAIC) in Shanghai, domestic chipmakers from Huawei Technologies to Biren Technology showcased new hardware designed to knit hundreds or thousands of chips together to act as one giant supercomputer. This is a direct response to the compute bottleneck that US export controls have imposed on China’s AI ambitions. The supernode is not just a technical concept; it is China’s strategic answer to the question of how to train frontier AI models without access to Nvidia’s most advanced hardware.
What a Supernode Actually Is — and Why It Matters
The term “supernode” is frequently misunderstood. As the South China Morning Post reports, “People think putting enough GPUs together makes a supernode, which is definitely not the case,” said Chen Daliang, CEO of Suanova, a Shanghai-based AI computing resources provider, speaking at WAIC. A true supernode requires not just a large number of chips but also a sophisticated architecture of high-speed interconnects, shared memory systems, and reliability engineering that allows the cluster to behave as a single coherent computing unit rather than a collection of individual processors.
This distinction matters because the challenge China faces is not simply acquiring more chips. It is building the system-level infrastructure that allows those chips to work together efficiently at scale. Morgan Stanley analysts, in a research note published during WAIC, noted that competition across China’s AI computing ecosystem has shifted away from stand-alone chip specifications and toward “interconnects, memory sharing, and system reliability.”
The supernode is the hardware embodiment of that shift. A chip that performs at 70% of Nvidia’s H100 in isolation may perform at 90% of a comparable Nvidia cluster when embedded in a well-designed supernode architecture, because the bottleneck in large-scale training is often inter-chip communication, not raw compute.
Huawei and Biren’s Competing Technical Approaches
The two most prominent supernode announcements at WAIC came from Huawei and Biren Technology, and they reflect meaningfully different technical philosophies. Huawei’s approach, embodied in its Atlas 950 SuperPoD unveiled at WAIC 2026, leverages the company’s existing Ascend chip ecosystem and its deep integration with Chinese cloud infrastructure. Huawei’s supernode strategy is essentially an extension of its existing chip and networking business, scaled up to meet the demands of trillion-parameter model training. The Atlas 950 SuperPoD can link up to 16 Ascend 910C nodes into a unified training cluster, with Huawei’s proprietary high-speed interconnect fabric managing the inter-chip communication.
Biren Technology has taken a more radical approach. The company has bet on light-based optical supernodes to break the 128-GPU cluster ceiling that has constrained conventional interconnect architectures. At WAIC, Biren unveiled a 1,024-GPU near-packaged optics (NPO) super node, using optical rather than electrical interconnects to achieve the bandwidth and latency characteristics needed to treat a thousand-GPU cluster as a single computing unit. If the technology matures, it could allow Chinese AI companies to build training clusters that approach the scale of the largest American facilities without requiring access to Nvidia’s NVLink interconnect technology.
The Bottleneck That Supernodes Must Solve — and Whether They Can
The urgency behind China’s supernode development is driven by a specific technical reality: training the next generation of frontier AI models requires not just more chips, but chips that can communicate with each other fast enough to function as a unified system. The largest American AI training runs use tens of thousands of Nvidia GPUs connected by NVLink and InfiniBand, achieving levels of inter-chip bandwidth that current Chinese alternatives cannot match. This communication bottleneck, not raw compute, is increasingly the binding constraint on the scale of training runs that Chinese AI companies can execute.
Supernodes are China’s answer to this problem, a way to achieve the effective scale of a massive training cluster using domestic hardware that, individually, may not match Nvidia’s specifications but, when properly interconnected, can approach the system-level performance needed for frontier model training. Whether this approach can close the gap with American infrastructure in time to matter for the current generation of model development is the central question hanging over China’s AI compute strategy. The answer will determine whether the next generation of Chinese frontier models can be trained domestically or whether they will continue to depend, in part, on hardware that Washington can restrict.