Huawei throws massive 1-million NPU gauntlet at Nvidia and AMD as it positions itself an alternative to US AI giants (www.techradar.com)

Huawei today unveiled what it calls the “world’s most powerful SuperPoDs and superclusters,” pitching a homegrown alternative to Nvidia- and AMD-led AI infrastructure. Its Atlas 950 SuperPoD will pack 8,192 Ascend NPUs (AI accelerators) and the Atlas 960 SuperPoD ramps that to 15,488 NPUs (≈89% more). Huawei says Atlas 950 systems, built on newly announced Ascend 950 chips, arrive Q1 2026, with Ascend 960/Atlas 960 in Q4 2027 and an Ascend 970 planned for Q4 2028. Aggregating SuperPoDs into superclusters, Huawei claims configurations exceeding ~520,000 NPUs (64 SuperPoDs) and topping one million NPUs (≥66 SuperPoDs), positioning them to outsize current large clusters like xAI’s Colossus. Technically notable is Huawei’s UnifiedBus interconnect, presented as an InfiniBand alternative with claims of 100× better optical reliability, >200 m range, and NPU-to-NPU latency of 2.1 ms (≈30% improvement). The announcement signals a strategic push to build a full-stack AI ecosystem—compute, network, software, storage and management—challenging Western suppliers and underscoring China-focused demand (enterprise and state). Important caveats remain: independent verification is absent, pricing, power efficiency and real-world throughput/scaleability aren’t disclosed, and competing hyperscalers (xAI Colossus 2, Oracle/OpenAI, Meta, AWS/Anthropic) are progressing on their own gigawatt-class clusters.