Huawei unveils Atlas 960E SuperPoD with 4,096 AI chips

At HUAWEI CONNECT 2026 in Shanghai on Sep 17, Huawei Rotating Chairman David Wang unveiled the Atlas 960E SuperPoD, which Huawei calls the industry’s first SuperPoD using near-packaged optics (NPO), according to the company’s release.

What the company says the box does: A single SuperPoD scales to 4,096 NPUs with unified memory addressing over UnifiedBus, delivering 8 EFLOPS at FP8 and 16 EFLOPS at FP4. Huawei says 5,500 Hi-ONE NPO engines replace about 48,000 traditional 800G optical modules, cutting power by more than 550 kW, doubling fault-free operating time, and reaching 99.8% system availability. Multiple SuperPoDs can form a SuperCluster that Huawei says can interconnect up to 512,000 NPUs (or up to one million with multi-rail topology).

Chip calendar (hedged): TechCrunch reports a Huawei spokesperson saying Ascend 960DT is “expected to be ready in Q1 2027,” ahead of a prior schedule — that date is not restated in the English SuperPoD release we checked.

Why it matters for markets and chips: Huawei is selling a domestic training-and-inference stack — Ascend silicon, SuperPoD packaging, and open-source CANN — as an alternative compute path for large models. That is silicon and power infrastructure, not a tip on any ticker.

Stake: Chinese labs and cloud buyers on Ascend; Nvidia and other GPU vendors in export-constrained markets; optical and liquid-cooling suppliers; investors tracking whether Huawei’s SuperPoD roadmap ships on the dates it names.

What’s next: A Huawei page that locks Ascend 960DT / 960PR dates; first customer SuperPoD deployments; how Atlas 960E (4,096 NPU) relates to earlier Atlas 960 (15,488-chip) plans.

— Nico Valdez