Huawei Unveils Ascend 960 SuperPoD with NPO for AI Infrastructure

Summary

At HUAWEI CONNECT 2026, Huawei announced the Ascend 960 SuperPoD, a system designed to address communication bottlenecks in very large AI clusters. Huawei said that in a 100,000-card cluster, inter-chip communication can consume more than 40% of training time under traditional architectures. The SuperPoD connects multiple computing nodes through high-speed interconnects so they can share unified memory and operate as one system. It supports up to 4,096 NPU cards, up to 8E FP8 computing power, and up to 1 petabyte of HBM capacity. Its distinguishing feature is Near-Packaged Optics, or NPO, which places optical components closer to the computing chips to reduce the distance traveled by high-speed electrical signals. Huawei’s Hi-ONE NPO product provides 7.2T of transmission capacity per engine; Huawei says about 5,500 Hi-ONE units can replace 48,000 800G optical modules and cut power consumption by more than 550 kW. The company also projects doubled mean time between failures and 99.8% availability. Huawei presented the system, together with UnifiedBus, as part of a broader shift from individual-chip competition toward integrated systems spanning compute, interconnects, memory, and storage. The company is also discussing clusters of up to one million cards as demand grows for continuously operating AI agents, autonomous driving, and AI-enabled devices.