Huawei used the eve of the World Artificial Intelligence Conference to make its most aggressive hardware statement since US sanctions cut it off from TSMC: the first public display of the Atlas 950 SuperPoD, staged at the Shanghai World Expo Exhibition Hall on July 16 and running as the conference's centerpiece exhibit through opening day.
What's on the floor
The demonstration configuration connects 1,024 Ascend NPUs across 16 cabinets — a fraction of the design's ceiling. At full build-out a single SuperPoD links 8,192 Ascend chips, which Huawei rates at 1 EFLOPS of FP8 compute and 2 EFLOPS at FP4, backed by 256TB of globally addressable memory. The system is billed as the industry's largest AI computing supernode.
The interconnect is the point
Where Nvidia stitches racks together with NVLink, Huawei's answer is Lingqu, also branded UnifiedBus — a proprietary protocol offering terabyte-class NPU-to-NPU bandwidth at a claimed 3-microsecond round-trip latency. Individual Ascend chips still trail Nvidia's silicon; the wager is that interconnect scale can make the cluster, not the chip, the unit of competition.
The Nvidia comparison
Huawei claims the Atlas 950 SuperPoD delivers 6.7x the computing power of Nvidia's NVL144 rack system — a company figure, not an independent benchmark. The roadmap stacks further: 64 SuperPoDs chain into the Atlas 950 SuperCluster carrying more than 520,000 Ascend 950DT chips, with availability from Q4 2026. Huawei emphasizes the entire stack is built without US components.
The audience
The display is aimed less at Silicon Valley than at Chinese hyperscalers weighing their options. Beijing has reportedly capped approvals for Nvidia H200 imports below 200,000 units and steered inference workloads toward domestic silicon — a policy that only works if machines like this one exist at scale. WAIC's 1,100-plus exhibitors and the state media apparatus covering Xi Jinping's keynote gave Huawei the largest possible stage to argue that they do.