Huawei Ascend 960 SuperPoD Debuts as World’s First NPO-Based AI Supernode
Huawei Technologies Co. unveiled the Huawei Ascend 960 SuperPoD on Thursday, September 17, 2026, at Huawei Connect 2026 in Shanghai, describing the system as the world’s first supernode built on near-package optics (NPO) technology. Deputy Chairman and Rotating Chairman Wang Tao used his opening keynote to frame the device as the cornerstone of the company’s strategy to accelerate training and inference for trillion-parameter artificial intelligence models.
The technical heart of the Huawei Ascend 960 SuperPoD, according to materials released at the conference, is what Huawei calls its UnifiedBus 2.0 interconnect combined with the Hi-ONE optical system. A single protocol treats every compute, memory, and storage component as an equal peer, enabling cross-server memory access across thousands of processors. Engineers described a near-copper, far-optical architecture in which short-range signals travel through conventional copper while longer-distance traffic is converted to light at the package level. The company said UnifiedBus and Hi-ONE together can link 4,000 processors simultaneously without the bottlenecks of traditional switched fabrics.
Huawei Ascend 960 SuperPoD Atlas 960 Configuration and Scale
At full configuration, the Atlas 960 SuperPoD integrates 15,488 Ascend chips spread across 220 cabinets and occupies roughly 2,200 square meters. Huawei said the system accelerates training and inference workloads by a factor of three to four compared with the prior generation, and can be aggregated with other Ascend 960 supernodes into clusters of up to 512,000 cards, or 1 million cards when using a multi-track topology. A companion arXiv paper published this month details a 256,000-node-class single AI computing framework that extends the architecture beyond a single SuperPoD, effectively replacing what would otherwise be more than 200,000 separate machines with one unified system.
Huawei also sharpened its near-term roadmap. The Ascend 960DT is scheduled for commercial launch in the first quarter of 2027, three quarters ahead of the original plan, while the Ascend 960PR is slated for the third quarter of 2027, one quarter earlier than previously announced. The company reiterated an annual cadence, confirming the Ascend 970 for 2028 and the Ascend 980 for 2029. “Supernodes are an inevitable choice for building ultra-large-scale data centers,” Wang Tao said during the keynote. “Huawei is set to begin mass production of the world’s first NPO product with an integrated light source.”
Huawei Ascend 960 SuperPoD Roadmap to 2029 With Annual Cadence
Wang framed the announcement as the realization of a “supernode plus cluster” innovation model that the company has been refining since the Atlas 950 generation. The Atlas 950 SuperPoD, which packs 8,192 Ascend NPUs, has been commercially deployed, while the more than 1,000 supernodes running the previous Ascend 910C chip have, in Huawei’s telling, reached an ecosystem tipping point. The Ascend 950 is already in commercial use at customer sites, providing the installed base into which the new systems will be folded.
The launch lands against a backdrop of tightening United States export controls that have restricted Chinese access to Nvidia’s most advanced accelerators. Huawei executives did not name Nvidia directly, but the timing and language of the announcement — emphasizing domestic mass production of optical interconnect, accelerated roadmaps, and cluster scales designed for frontier model training — point clearly at efforts to reduce reliance on restricted foreign silicon. Industry analysts at Counterpoint Research and Bernstein have previously estimated that Chinese hyperscalers are absorbing every domestic Ascend shipment Huawei can produce, and the move to NPO is expected to extend that supply advantage into the optical networking layer as well. The Huawei Ascend 960 SuperPoD will be measured against this competition, and the Huawei Ascend 960 SuperPoD launch closes a quarter of intense activity across the Chinese AI hardware stack.
For the broader AI infrastructure market, the Huawei Ascend 960 SuperPoD introduces a new variable into a contest previously dominated by Nvidia’s NVL rack-scale designs and AMD’s Helios roadmap. By collapsing the boundary between server, rack, and data center through UnifiedBus 2.0 and Hi-ONE, Huawei is betting that supernode architectures will define the next generation of training clusters regardless of the underlying chip vendor. If mass production begins on schedule and the early roadmap targets hold, the Huawei Ascend 960 SuperPoD will move from keynote slide to shippable product in the months ahead, setting the pace for the 2027 AI hardware cycle and sharpening competition in trillion-parameter model infrastructure built on the Huawei Ascend 960 SuperPoD.
Beyond the silicon, the Huawei Ascend 960 SuperPoD shift reframes how Chinese hyperscalers will think about cluster composition over the next twenty-four months. JD Cloud and Moore Threads have separately announced plans to build 100,000-GPU computing clusters on domestic silicon, and the Huawei Ascend 960 SuperPoD provides the first commercially deployed reference architecture at that order of magnitude for an Nvidia alternative. With Ascend 950 already in commercial use and Ascend 910C supernodes numbering more than 1,000 in deployment, the Huawei Ascend 960 SuperPoD enters a maturing ecosystem rather than an empty one. As mass production begins and the 2027 roadmap rolls forward, the Huawei Ascend 960 SuperPoD will serve as the most visible signal of how China intends to scale frontier AI training under continued US export controls.
Source: ChinaDaily