Huawei's new AI computing architecture moves optical integration inside the node to scale processing power. The Ascend 960 supernode links 4,096 chips to reach 8 EFLOPS FP8 performance while replacing 48,000 800G optical modules with 5,500 Hi-ONE near-package optics engines. This design reduces power consumption by more than 550 kW compared to previous layouts.

Rotating chairman Wang Tao introduced the system at the Huawei Connect event on September 17 for shipment in the third quarter of 2027. The hardware targets a 99.8% availability rate and doubled fault-free runtime to challenge Nvidia's scale in the AI market. Third party verification of the supernode's specifications is not currently available.