At HUAWEI CONNECT 2026 in Shanghai, Huawei Deputy Chairman and Rotating Chairman David Wang outlined the company's strategic focus on building AI infrastructure to meet the demands of an intelligent world. With foundation model parameters projected to exceed 100 trillion by 2030 and daily inference tokens in China expected to reach quintillions, the need for massive, efficient computing power is paramount. Huawei's response centers on SuperPoDs and SuperClusters, aiming to provide a solid silicon foundation.
Central to this strategy is the newly unveiled Atlas 960E SuperPoD, the industry's first to use near-packaged optics (NPO). A single Atlas 960E SuperPoD can scale to 4,096 NPUs, delivering 8 EFLOPS of FP8 compute and up to 1 petabyte of HBM capacity. It leverages Huawei's Hi-ONE optical interconnect, which achieves 7.2 Tbit/s per engine and eliminates the need for 48,000 traditional 800G optical modules, cutting power consumption by over 550 kilowatts and doubling fault-free operating time to achieve 99.8% availability. This breakthrough addresses the critical bottleneck of interconnect overhead in large-scale AI clusters, where communications can account for over 40% of training time.
Huawei also upgraded its TaiShan 950 SuperPoD, now supporting up to 4,096 nodes with a unified memory pool of 256 TB, and launched the OceanStor M900, a context memory storage system with petabyte-scale KV caching for agentic inference. Together, these form an agentic SuperCluster that uses UnifiedBus to unify interconnect protocols, enabling peer-to-peer connectivity among Ascend SuperPoDs, Kunpeng SuperPoDs, and KV cache clusters. The SuperCluster can scale to 512,000 NPUs, and with multi-rail topology, up to one million NPUs, providing the scale needed for 10-trillion-parameter models.
The implications are significant. As AI models grow, traditional server architectures struggle to keep pace. Huawei's SuperPoD approach, which tightly couples nodes to function as a single logical computer, can deliver a 2.75x increase in Model FLOPs Utilization compared to clusters built from 8-NPU servers, according to Huawei's Markov Lab simulations. This efficiency gain is crucial for training and inference at scale, potentially lowering costs and accelerating AI innovation.
Huawei's commitment to open ecosystems further amplifies its impact. The Kunpeng ecosystem has attracted over 4.16 million developers and 7,200 partners, while CANN has transitioned to community-driven open-source development, with external developers now comprising 61% of the total. Ascend supports over 90 leading open-source projects, including PyTorch, and is the first official Chinese compute platform on PyTorch's website. This openness fosters collaboration and ensures that advancements benefit the broader AI community.
For more information, visit HUAWEI CONNECT 2026.
