What is a supernode, and why does it matter for the China-US tech rivalry?

As the size of artificial intelligence models expands beyond 1 trillion parameters, a new concept is dominating the computing landscape: “supernode”. At this year’s World Artificial Intelligence Conference (WAIC), China’s top AI summit, 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, which they called a supernode. The focus on supernodes at WAIC reflected a shift in competition across China’s AI computing ecosystem – away from stand-alone chip specifications and towards “interconnects, memory sharing, and system reliability”, Morgan Stanley analysts said in a research note on Tuesday. But can supernodes immediately address China’s AI computing power bottleneck? What is a supernode? “People think putting enough GPUs [graphics processing units] together makes a supernode, which is definitely not the case,” said Chen Daliang, CEO of Suanova, a Shanghai-based AI computing resources provider and a subsidiary of Hong Kong-listed Yeebo Technology. Instead, supernodes deeply integrate computation, storage, networking, management, and software frameworks so that the whole is greater than the sum of its components. Properly built, supernodes can boost computing power utilisation rate by more than 50 per cent while doubling network communication bandwidth. To achieve scale, the industry is pursuing two interconnect road maps: electrical and optical. The electrical path increasingly adopts orthogonal architectures, which alter the spatial geometry of the system to shorten signal pathways. Optical solutions include Near-Packaged Optics (NPO), which moves the optical engine closer to the computing chip on the circuit board to reduce power use and electrical transmission distances while maintaining high bandwidth. For now, most companies remain in the prototype stage in terms of optical interconnects. Domestic progress in supernodes At this year’s WAIC, which concluded on July 20 in Shanghai, Huawei made the first public display of its Ascend Atlas 950 SuperPoD, seen as a peak of domestic compute systems. Using 64-card cabinets as its base unit, it can support up to 8,192 cards through the proprietary “Lingqu” high-speed network. Telecoms equipment maker ZTE also showcased a new supernode solution capable of supporting up to 16,384 GPUs and compatible with several domestic chips. AI infrastructure maker Sugon exhibited its 100,000-card “Sugon 8000” supercluster powered entirely by domestic chips. Shanghai-based Biren Technology unveiled a 1,024-card NPO supernode. Other key players including MetaX and Enflame also unveiled progress on prototypes. Suanova’s Chen said the primary buyers were currently state-level data centre operators rather than private companies like Alibaba Group Holding, owner of the South China Morning Post. Alibaba has been working on its proprietary supernode solutions based on the Zhenwu M890 chip, its most powerful AI processor to date. The company announced on Thursday that supernodes based on the flagship chip had been adopted by its newly released Qwen3.8 model, marking China’s first deployment of a model exceeding 2 trillion parameters on a domestic supernode architecture. However, a Tencent executive said it expected to deploy NPO supernodes by the fourth quarter of 2026, becoming the first Chinese tech company to have disclosed a clear timeline for adoption, according to Chinese media reports. The rise of supernodes reflects the increasingly modular nature of China’s AI server ecosystem. “Accelerator vendors provide the computing platform and software stack, while ODMs [original design manufacturers] integrate servers, interconnect, cooling and power infrastructure into rack- or cluster-level solutions,” Morgan Stanley analysts said. Can it address the computing power challenge? By pooling the power of individual chips to train massive AI models, supernodes are expected to help China bypass US export controls, offering a vital lifeline amid difficulty in accessing advanced chips from Nvidia. But the reality is more complicated as supernodes are difficult to build and scale. Based on the Suanova’s experience deploying supernode systems since last year, Chen said the initial design cost alone could exceed 100 million yuan (US$14.8 million), with a single prototype costing up to 30 million yuan. Optimising the network-scheduling platform took about 18 months, by which time the design risked becoming obsolete amid rapid product cycles. Because these high-density systems consumed massive power, they required supporting infrastructure such as data centre power and cooling systems to upgrade, which was becoming “mandatory” rather than “optional”, according to a recent report by Huatai Securities. They included 800V high-voltage direct current (HVDC) power to reduce loss, and liquid cooling systems to manage soaring cabinet temperatures, the report added. Not all workloads need such a big cluster. While hyper-local weather forecasting might require 2,000 cards, a hospital running multi-modal medical imaging might only need four for a specific analysis of medical scans, or 96 to train internal medical models, Chen said. “Bigger isn’t always better,” he said.

What is a supernode, and why does it matter for the China-US tech rivalry?

.css-1c6uqr6{color:inherit;font-weight:inherit;font-size:inherit;font-family:inherit;line-height:inherit;overflow-wrap:break-word;}Tech firms are racing to build colossal clusters for next-gen AI, but cost, power and scale remain challenges

What is a supernode?

Tech firms are racing to build colossal clusters for next-gen AI, but cost, power and scale remain challenges