Huawei Atlas 950 SuperCluster unveiled with 524 ExaFLOPS AI performance

Huawei has introduced the Atlas 950 SuperCluster at Huawei Connect 2025, delivering exascale performance with 524 FP8 ExaFLOPS for AI training and 1 FP4 ZettaFLOPS for inference. This system is positioned to compete directly with Nvidia's Rubin-based platforms, expected in late 2026. The Atlas 950 integrates 524,288 Ascend 950DT neural processing units across over 10,240 cabinets.

The architecture comprises 64 Atlas SuperPoDs, each containing 8,192 Ascend 950DT chips, offering 20 times more processing units than the previous Atlas 900 A3 systems. The new architecture supports both RoCE and Huawei's proprietary UBoE (UnifiedBus over Ethernet) protocols, and features all-optical interconnect with 16 PB/s bandwidth and 2.1-microsecond latency.

The Atlas 950 SuperCluster surpasses Oracle's OCI Supercluster, which uses 131,072 B200 GPUs to deliver 2.4 FP4 ZettaFLOPS. Atlas 950 also exceeds xAI's Colossus cluster, offering 2.5 times more NPUs and 1.3 times greater compute capacity. Each Atlas 950 SuperPoD provides 8 FP8 ExaFLOPS, compared to 1.2 FP8 ExaFLOPS for Nvidia's Vera Rubin NVL144. Huawei has designed the SuperCluster to support AI models ranging from hundreds of billions to tens of trillions of parameters.

Each SuperPoD covers about 1,000 square meters, which is similar to the area of two basketball courts, and requires 160 cabinets. Altogether, the SuperCluster spans 64,000 square meters, or about 150 basketball courts or nine soccer fields. Rather than using more powerful chips like Nvidia, Huawei relies on a larger number of accelerators, so the cluster needs more space for power, cooling, and support systems.

Huawei expects to launch the Atlas 950 SuperCluster in late 2026, with the 960 SuperCluster coming a year later. The 960 will include more than a million NPUs and offer two to four ZettaFLOPS of computing power. Because Huawei’s chips are less powerful on their own, the company is building larger systems overall. These SuperClusters are designed for AI companies that need massive computing power for their next-generation models.