Huawei’s Atlas 960E SuperPoD: What Was Announced at Connect 2026

At Huawei Connect 2026 in Shanghai, Huawei unveiled the Atlas 960E SuperPoD, an Ascend 960-based AI system built around connecting many processors through a shared interconnect. The announcement sets out a system design and performance targets; it does not establish that a 4,096-NPU installation is already in service. 136

How the SuperPoD is designed to scale

Huawei describes the Atlas 960E as a fully liquid-cooled system with an orthogonal architecture. It combines near-packaged optics (NPO) with Lingqu UnifiedBus and Hi-ONE, Huawei’s optical interconnect engine. UnifiedBus enables unified memory addressing across the SuperPoD; Huawei says that design lets one system scale to 4,096 NPUs, or computing cards. 1319

The emphasis is on moving data between processors as well as supplying processing power. Huawei says the Atlas 960E integrates 5,500 Hi-ONE units in place of the 48,000 800G optical modules it estimates a more conventional connection scheme would require. Those component counts and the implied efficiency benefit are Huawei’s comparisons, not independently verified measurements here. 13

Performance claims—and a latency caveat

Huawei says the system can support training and inference for models with up to 10 trillion parameters. For a 10-trillion-parameter training scenario, it claims model-FLOPS utilization 1.1 times that of the previous generation; for inference, it claims 2.5 times the prior generation’s single-card throughput. These are different metrics, not interchangeable claims about overall training speed. 106

A figure of roughly 2 microseconds round-trip latency also appears in secondary reporting about the interconnect. The provided Huawei announcements do not substantiate that number, so it should be treated as a reported claim rather than a verified benchmark. 2513

When is Ascend 960 due?

Huawei moved its planned availability for the Ascend 960DT to Q1 2027, three quarters earlier than its previous roadmap, and the Ascend 960PR to Q3 2027, one quarter earlier. Huawei announced the liquid-cooled Atlas 960E alongside that accelerated chip roadmap, but unveiling the system is not the same as confirming its volume-shipment date. 113

The wider strategy extends beyond one SuperPoD: Huawei has outlined a UnifiedBus-based chip portfolio and a path toward SuperClusters with as many as one million processors. That figure is a roadmap ambition, not an operating-cluster count. 311

What about existing customers? One report attributes to Huawei a claim that more than 1,000 Atlas SuperPoD systems had been deployed for about 370 customers, in the context of the earlier Atlas 950. It does not establish shipments or customers for the newly announced Atlas 960E, nor does it name those customers. 22