Huawei Unveils Ascend 960 SuperPoD: Industry-First NPO with 4,096 Cards and 8 EFLOPS, Ascend 960 Chip Pulled Forward Three Quarters

On September 17, 2026, Huawei used the opening day of its annual HUAWEI CONNECT 2026 event in Shanghai to announce the Ascend 960 SuperPoD, a new AI computing system built around its next-generation Ascend 960 accelerator. A single SuperPoD bundles 4,096 cards to deliver up to 8 EFLOPS of FP8 compute and 1 PB of HBM capacity, and Huawei says it is the first system in the industry to adopt Near-Packaged Optics (NPO), which places optical modules right next to the processors. The company also said development of the Ascend 960 chip is running ahead of schedule, with the training-oriented 960DT now due in the first quarter of 2027.

What SuperPoDs Are For, and How Many Are Deployed

A SuperPoD ties a large number of NPUs (AI processors) together over a high-speed interconnect so they behave like one enormous computer. In his keynote, Deputy Chairman and Rotating Chairman Wang Tao explained that clusters of around 100,000 cards are becoming standard for training models with 10 trillion parameters, yet in conventional server designs intra-cluster communication eats up more than 40 percent of training time, dragging down model FLOPS utilization (MFU). According to Huawei's internal simulations, a 100,000-card cluster built from 4,096-card SuperPoDs achieves 2.75 times the MFU of a cluster of the same size built from 8-card servers.

On the deployment side, Huawei said more than 1,000 SuperPoDs based on the two-generations-old Ascend 910C are already in operation, and that the previous-generation Ascend 950 SuperPoD has entered full commercial use. The Ascend 960 SuperPoD carries that line into the next generation.

Optics Next to the Chip: Hi-ONE Cuts Power by 550 kW

The technical centerpiece of the announcement is Hi-ONE, Huawei's in-house optical interconnect product. With NPO, the optical modules that used to plug into a switch's front panel are mounted close to the processor instead. Because electrical signals no longer travel long distances across the board, the retimer chips that clean up those signals can be removed, reducing both power consumption and latency. Hi-ONE offers 7.2 Tbps of transmission capacity per engine, and Huawei claims it is the only mass-produced NPO product with a built-in light source.

In the Ascend 960 SuperPoD, 5,500 Hi-ONE units replace the 48,000 800G optical modules that would otherwise be required, cutting power consumption by more than 550 kW, according to the company. Huawei also says the system's mean time between failures doubles and availability reaches 99.8 percent. The company has proposed an NPO standardization project at the OIF, the optical interconnect standards body, in an effort to grow the ecosystem around the technology.

One Ascend Generation per Year, with 970 and 980 Already Announced

Progress on the chip itself is another highlight. Wang said the Ascend 960 doubles the performance of its predecessor and that development has exceeded expectations. The Ascend 960DT, aimed at training and decode-stage inference, will be ready in the first quarter of 2027, three quarters ahead of the original plan. The Ascend 960PR, aimed at prefill-stage inference and recommendation workloads, will follow in the third quarter of 2027, one quarter early.

As products, the air-cooled SuperPoD is expected to reach the market in the second quarter of 2027, and the liquid-cooled Atlas 960 SuperPoD in the third quarter of 2027. Huawei said it will keep updating Ascend on a one-generation-per-year cadence, with the Ascend 970 in 2028 and the Ascend 980 in 2029, each doubling compute while substantially raising memory bandwidth, memory capacity, and interconnect bandwidth.

Kunpeng SuperPoD and Dedicated KV-Cache Storage

Beyond Ascend for AI, Huawei also refreshed the SuperPoD built on its general-purpose Kunpeng CPUs. Using an all-optical network based on its proprietary Lingqu UnifiedBus interconnect, it links up to 4,096 nodes and forms a unified memory pool of 256 TB. Huawei says launching sandboxes for AI agents at a scale of 100,000 is 30 times faster than on conventional servers, and that large-scale vector search performance doubles.

For inference systems, the company introduced OceanStor M900, a dedicated storage tier that holds PB-scale KV cache (the data generative AI uses to retain conversational context). NPUs can reach it in a single hop over UnifiedBus, and Huawei says a combination of mixed media and algorithmic tuning extends SSD write endurance 16-fold. The company also presented an Agentic SuperPoD Cluster that connects Ascend SuperPoDs, Kunpeng SuperPoDs, and KV-cache clusters as peers over the same UnifiedBus, scaling to 512,000 cards, or up to 1 million cards when combined with multi-rail topology.

State of the Developer Ecosystem

On the software side, Huawei said CANN, its development platform for Ascend, is now run as an ongoing open-source community, and that external developers account for 61 percent of contributors, surpassing internal developers for the first time. Ascend has been registered as a compute platform installable directly from the official PyTorch site, the first from China, and supports more than 90 major communities including PyTorch, Triton, and vLLM. For Kunpeng, Huawei reported more than 4.16 million developers worldwide and over 7,200 ecosystem partners.

It is worth noting that every performance and power-saving figure presented here comes from Huawei itself; no third-party verification has been published yet. The previous-generation Ascend 950 has been described as trailing NVIDIA GPUs in performance, so how close the Ascend 960 actually gets will only become clear once it ships in 2027.

Summary

At HUAWEI CONNECT 2026, Huawei announced the Ascend 960 SuperPoD with 4,096 cards, 8 EFLOPS, and 1 PB of HBM, and said its Hi-ONE NPO optical interconnect cuts power consumption by more than 550 kW. The Ascend 960 chip has been pulled forward, with the 960DT ready in the first quarter of 2027 and the 960PR in the third quarter, while the air-cooled SuperPoD arrives in the second quarter and the liquid-cooled version in the third. With access to NVIDIA products restricted in China, the announcement makes Huawei's strategy of winning on whole-system design clearer than ever.