At HUAWEI CONNECT 2026, Huawei introduced the Ascend 960 SuperPoD, leveraging innovative Near-Packaged Optics (NPO) technology to significantly enhance AI infrastructure performance by tightly integrating chip communication and memory within massive AI clusters.
- Ascend 960 SuperPoD supports up to 4,096 NPU cards and 1PB memory.
- NPO reduces power use by 550kW versus traditional optical modules.
- System availability reaches 99.8% with doubled mean time between failures.
What happened
Huawei unveiled its Ascend 960 SuperPoD at HUAWEI CONNECT 2026, positioning it as a fundamental building block for next-generation AI infrastructure. This system integrates up to 4,096 neural processing unit (NPU) cards and offers 8 exaFLOPS (FP8) computing power alongside 1 petabyte of high-bandwidth memory (HBM).
A standout innovation is the use of Near-Packaged Optics (NPO) technology, which relocates optical components closer to computing chips to overcome bandwidth, latency, and power inefficiencies inherent in traditional copper and optical module interconnects. Huawei’s Hi-ONE NPO product replaces tens of thousands of 800G optical modules, significantly reducing energy consumption and enabling much faster data exchange within AI clusters.
Why it matters
AI workloads are rapidly scaling, with models reaching the 10-trillion-parameter level and beyond, stressing traditional multi-server setups where chip communication can take over 40% of training time. Huawei’s approach with the Ascend 960 SuperPoD and NPO technology tackles these bottlenecks by unifying memory and computing resources, creating a highly efficient and integrated AI cluster architecture.
By improving data transfer rates, reducing power demands, and increasing system reliability—achieving 99.8% availability and doubling mean time between failures—Huawei is not just advancing chip performance but also pioneering a holistic infrastructure shift. This model addresses how AI computing power, memory, interconnects, and storage collaborate to handle future AI demands across devices and real-time applications.
What to watch next
Huawei’s development signals a new era of AI infrastructure where the integration of chips, optical interconnects, and storage defines performance more than standalone chip advancements. Monitoring adoption of the Ascend 960 SuperPoD and NPO tech in AI datacenters will reveal whether Huawei’s unified system approach gains traction against competitors focusing mainly on raw chip power.
Additionally, the expansion plans for AI clusters potentially reaching up to one million cards suggest scaling challenges ahead, including managing power efficiency, system uptime, and real-time computing. Future announcements around ecosystem partnerships, software optimization, and deployment in autonomous driving, IoT, and AI-driven devices will be key indicators of Huawei’s infrastructure strategy impact.