Chinese AI accelerator and infrastructure vendors, including Huawei and Biren, are launching rack-scale supernodes with 64 to over 1,000 cards to offset individual chip performance gaps. These supernodes feature a high XPU-to-switch ratio of 2:1, significantly increasing demand for PCIe switches to manage complex I/O configurations. Moonshot AI's Kimi K3 model, with 2.8 trillion parameters and 1M-token context, further validates infrastructure demand, requiring about 1.4TB of HBM capacity and placing a premium on memory bandwidth.
China's AI infrastructure buildout is prioritizing system-level integration and driving specific demand for PCIe switches and high-bandwidth memory.