Huawei used its annual HUAWEI CONNECT keynote to unveil the Ascend 960 SuperPoD, which the company calls the industry’s first SuperPoD built on near-packaged optics.
The system links up to 4,096 NPU cards, delivering up to 8E FP8 compute and 1PB of HBM capacity. Huawei argues the architecture matters more than the headline numbers: in a 100,000-card cluster, chip-to-chip communication can eat more than 40% of training time under conventional server designs.
Its answer is NPO, or near-packaged optics, which moves optical components closer to the compute silicon so high-speed electrical signals travel shorter distances. Huawei’s Hi-ONE engine carries 7.2T of transmission capacity. Roughly 5,500 Hi-ONE units replace the 48,000 800G pluggable optical modules a comparable build would need, cutting power consumption by more than 550kW. Mean time between failures doubles, Huawei says, and availability reaches 99.8%.
Rotating chairman Wang Tao framed the launch around the agentic era, where models at the 10-trillion-parameter scale outgrow the wiring that connects them.
The pitch lands in a market where export controls have pushed Chinese labs toward domestic silicon. Huawei’s argument is that interconnect, memory and system design – not individual chips – decide who can train the largest models.
