Huawei Unveils Ascend 960 AI Chip and SuperPoD Cluster to Challenge NVIDIA Dominance

At the opening of Huawei Connect 2026 in Shanghai on September 17, 2026, Huawei Technologies officially unveiled its next-generation Ascend 960 artificial intelligence processor alongside the Ascend 960 SuperPoD computing cluster. Presented by rotating chairman Wang Tao, the new AI platform introduces an upgraded UnifiedBus interconnect and near-package optics (NPO) architecture designed to scale up domestic frontier AI model training and reduce reliance on NVIDIA silicon.

Ascend 960 Architecture and SuperPoD Scale-Up

Addressing thousands of developers and enterprise partners at the Shanghai World Expo Exhibition and Convention Center, Huawei outlined its accelerated hardware roadmap for AI computing. The Ascend 960 NPU is engineered to deliver substantial throughput improvements for large language model pre-training and high-concurrency inference, packing massive high-bandwidth memory capacity and high-density tensor cores. To circumvent overseas supply chain bottlenecks, Huawei emphasized a cluster-level scale-out paradigm through its Atlas and Ascend 960 SuperPoD architecture, linking tens of thousands of accelerator cards with unified memory addressing.

Technical Architecture & Compute Capabilities

• Compute Throughput: Targets 1,000 TFLOPS of FP16 tensor precision alongside specialized INT8 and FP8 mixed-precision matrix execution engines.
• Memory Subsystem: Features up to 288GB of integrated high-bandwidth memory per accelerator package.
• UnifiedBus Interconnect: Upgraded proprietary interconnect protocol delivering low-latency peer-to-peer communications across distributed nodes.
• Near-Package Optics (NPO): Integrates optical interconnect technology to boost rack-to-rack bandwidth while containing electrical power dissipation.
• Commercial Availability: Sampling to foundational model builders ahead of full enterprise commercial deployment targeted for 2027.

Challenging the Global AI Accelerator Landscape

The announcement represents Huawei most aggressive technical maneuver to date to provide a self-sufficient alternative to Western accelerator platforms. Amid tightened export restrictions, Huawei has pivoted toward software-hardware co-design within its MindSpore and CANN ecosystems. By optimizing distributed parallel algorithms across the Ascend 960 SuperPoD, the company aims to enable Chinese hyperscalers, telecommunication operators, and enterprise research labs to train frontier multimodal foundation models at sovereign scale.

Product Profile Sheet