Analysis:ChinaBetsonChipClusterstoSurviveU.S.Sanctions

Huawei’s Ascend 384 supernode is displayed at the 2025 World Artificial Intelligence Conference in Shanghai on July 28, 2025. Photo: VCG

Facing U.S. sanctions blocking access to advanced semiconductors, Chinese tech companies are increasingly relying on supernode technology that can integrate clusters of lower-performance chips into servers to achieve similar computing capacities. But systematic complexity, high costs and insufficient demand cloud their large-scale application.

Since mid-2025, many Chinese artificial intelligence (AI) chip designers, server manufacturers and cloud services providers have rolled out their own supernode products amid a surge in demand for training large language models (LLMs) that have been applied in a wide range of industries from education and health care to finance and marketing.

For example, Huawei Technologies Co. Ltd. in June launched the CloudMatrix 384, an AI computing system built with 384 Ascend 910C chips that is capable of delivering almost double the computing power of Nvidia Corp.’s GB200 NVL72 system. Separately, each Ascend 910C chip provides only one-third of the performance of Nvidia’s advanced Blackwell architecture chips.

In September, Alibaba Group Holding Ltd. unveiled the Panjiu AI server powered by a cluster of 128 self-developed chips. And in November, Baidu Inc. released similar products and announced plans to launch a computing system that can house 256 chips in the first half of 2026.

The endeavors reflect a change in China’s AI development logic: with access to cutting-edge semiconductors and high-bandwidth memory (HBM) chips restricted, domestic manufacturers are prioritizing system-level connections and communication efficiency to sustain the training of AI models. A supernode architecture integrates a large number of GPUs or AI chips within a server through high-speed interconnects in a setup that can significantly reduce communication latency and energy consumption.

While Nvidia currently utilizes its NVLink protocol to connect up to 72 GPUs in a server, Chinese companies are pushing for significantly higher numbers. Industry insiders said that while Nvidia dominates through a vertically integrated ecosystem of chips, protocols and software, China’s fragmented computing power ecosystem requires deeper coordination among providers of chips, servers and system software.

However, the push for massive supernodes has drawn skepticism regarding its technical viability and market demand. A source at a domestic AI chip manufacturer said that due to restrictions on advanced manufacturing processes and HBM purchases, Chinese firms are forced to stack more chips to bridge the performance gap with global leaders. While understandable as a technical workaround, this approach dramatically increases system complexity.

Expanding a cluster to hundreds of chips multiplies the requirements for heat dissipation, power supply, communication and system stability, making the engineering challenge immense, the source said. These large-scale systems, often costing tens of millions of yuan, are primarily targeted at specific inference scenarios for big state-owned enterprises rather than the broader commercial market.

In the competitive commercial market, efficiency and unit cost matter more than how many chips a cluster has. A person working at IT infrastructure provider IEIT Systems Co. Ltd. said that while its current servers can achieve an inference cost of approximately 1 yuan (14 U.S. cents) per 1 million tokens, mass adoption of large models as a basic utility requires costs to drop by an order of magnitude annually.

Several industry experts argued that supernodes built with approximately 64 chips are sufficient to cover the vast majority of current inference needs, with marginal benefits declining rapidly beyond that point, adding that the focus should be on reliability rather than scale.

A representative from a computing power firm said that the market currently does not need supernodes with hundreds of chips, but rather systems that are stable and capable of long-term operation at existing scales, warning that excessive expansion could lead to unnecessary resource consumption.

Contact editor Ding Yi (yiding@caixin.com)