Huawei’s Atlas 960E SuperPoD Puts Optical Links at the Center of AI Scale

Huawei unveiled the Atlas 960E SuperPoD at HUAWEI CONNECT 2026 in Shanghai on September 17, describing it as the first SuperPoD to use near-packaged optics (NPO). The announcement was about connecting thousands of AI processors into one system for large-model training and inference—not introducing a single faster NPU. 169

How UnifiedBus and Hi-ONE work together

UnifiedBus is the system interconnect; Hi-ONE supplies its near-packaged optical links. Placing optical engines close to processor packages shortens the electrical portion of the connection. Huawei rates each Hi-ONE engine at 7.2 Tbit/s and says it incorporates a light source within the product. That capacity describes an optical engine, not the measured throughput of an entire SuperPoD. 6122021

The design is intended to ease the data-movement challenge of connecting many NPUs. Huawei says a single Atlas 960E can scale to 4,096 NPUs, with peak compute ratings of 8 EFLOPS at FP8 and 16 EFLOPS at FP4. FP8 and FP4 are different precision modes, so the two figures should not be added together or read as real-world application benchmarks. Huawei also specifies up to 1 petabyte of high-bandwidth memory (HBM) per pod. 1315

Why the design matters for large models and long context

Training and inference across thousands of processors require data to move between them, while model parameters and working data need memory capacity. Huawei positions the Atlas 960E for models as large as 10 trillion parameters; the combination of optical connectivity and a large HBM pool is its proposed infrastructure for such workloads, including long-context use. The published 7.2 Tbit/s engine rating does not establish end-to-end long-context speed or latency. 16912

What Huawei claims about power and reliability

Huawei says a configuration using about 5,500 Hi-ONE engines avoids the need for roughly 48,000 conventional 800G optical modules. It claims the comparison would save more than 550 kW, double fault-free operating time and bring system availability to 99.8%. These are Huawei’s design comparisons and reliability claims, not independently verified results from a deployed Atlas 960E. 52122

Not the same as Huawei’s other announcements

The TaiShan 950 SuperPoD is an upgraded general-purpose computing system, while the Ascend roadmap concerns processors and CANN’s open-source development concerns AI software. OceanStor M900 is a memory-context storage system. They belong to the wider infrastructure story, but none is another name for the NPO-connected Atlas 960E AI compute pod. 161520