Huawei unveils Ascend 960 SuperPoD with NPO technology

Summary

Huawei unveiled the Ascend 960 SuperPoD at the HUAWEI CONNECT 2026 event. This 4,096-NPU system, built around Near-Package Optics (NPO) connections, claims to replace 48,000 traditional optical modules while reducing power consumption by over 550 kW.

Why it matters

As AI models scale toward 10 trillion parameters, raw FLOPS are increasingly overshadowed by interconnect bandwidth and power efficiency as the true bottlenecks. Huawei's move toward system-level architecture—featuring unified memory, NPO, and clusters scaling up to one million cards—signals that the AI infrastructure race is shifting from single-chip performance to full-stack integration. This raises industry expectations regarding what 'competitive' means beyond the chip itself.

Highlights

• The Hi-ONE NPO engine delivers 7.2T performance per unit; approximately 5,500 units replace 48,000 800G optical modules
• Huawei claims that in traditional architectures, over 40% of training time in a 100,000-card cluster is lost to inter-chip communication
• The system targets 99.8% availability and double the mean time between failures compared to previous designs

Critical take

All performance metrics cited in this article—8E FP8, 1 PB HBM, 550 kW savings, and 99.8% availability—originate directly from Huawei's own keynote. There is no independent benchmarking or third-party verification. The claim of being the 'industry's first' NPO SuperPoD is a marketing assertion. Furthermore, the article fails to contextualize the actual advantage by comparing it against rival interconnect approaches, such as other vendors' co-packaged optical solutions.