Ascend 960 to Launch in Q1 Next Year! Huawei Announces World's First NPO-Based Ascend 960 Super-Node for Q3 2027, Plans Annual Chip Evolution

Chinese tech giant Huawei today (17th) announced the Ascend 960 super-node at the 2026 Huawei Connect conference. Scheduled for launch in Q3 2027, Huawei's Vice Chairman and Rotating Chairman Wang Tao described it as the world's first super-node to adopt Near-Packaged Optics (NPO) technology, capable of supporting large-scale training and inference for 10-trillion-parameter AI models, further enhancing high-speed interconnect capabilities for ultra-large-scale AI infrastructure.

According to the announcement, the Ascend 960 features a single super-node scale of up to 4,096 cards, leveraging the Lingqu (UnifiedBus) architecture and Hi-ONE optical engine to achieve high-speed interconnectivity across all nodes, with RTT latency as low as 2 microseconds—one of the largest-scale super-nodes in the industry today. The NPO solution places the optical engine close to the main chip package, reducing bottlenecks from traditional pluggable optical modules and long-distance electrical interconnects. Combined with unified memory addressing, this enables communication across tens of thousands of chips to achieve near-linear bandwidth and latency, allowing them to work in concert like a single computer and improving MFU computational efficiency.

In terms of product roadmap, the Ascend 960DT is scheduled for launch in Q1 2027, with the Ascend 960PR and Atlas 960 liquid-cooled super-node planned for release in Q3 2027. Huawei will continue its annual chip evolution, with Ascend 970 slated for 2028 and Ascend 980 expected in 2029, doubling compute scale, bandwidth capacity, and near-memory storage bandwidth.

Wang Tao also revealed that Huawei has developed 11 series of chips, including Ascend and Kunpeng, centered around super-nodes and clusters, covering intelligent computing, general-purpose computing, interconnectivity, and storage. The next-generation NPO optical interconnect products are about to enter mass production.

In current deployments, over 1,000 sets of the Ascend 910C super-node have already been deployed, and the Ascend 950 super-node has begun large-scale commercial use for large-model training and high-concurrency inference in industries such as finance, government, and manufacturing.

Wang Tao stated that Huawei's AI strategy centers on computing power, with a commitment to hardware monetization. Through system architecture innovation using super-nodes and clusters, Huawei aims to build a robust domestic computing foundation in China. Simultaneously, it is building an open and open-source ecosystem, supporting native training of mainstream large models on Ascend, with the Pangu large model focusing on intelligentizing Huawei's own products, and offering flexible on-premise and cloud-based computing solutions. On the edge side, Huawei will promote diverse computing capabilities—from large to small—covering AI on devices, in vehicles, and lightweight intelligence for IoT. On the network side, it is building next-generation communications centered on 'using computing,' delivering intelligence to enterprises, homes, and individuals.

Analysts believe that as China's domestic large models evolve toward 10-trillion parameters, long-context understanding, and agent-based architectures, Huawei's use of optical interconnects, unified protocols, and annual chip iterations can reduce the cost of large-scale training and inference. These can be integrated with open-source ecosystems and industry-specific cluster solutions, though actual performance, supply chain stability, and software migration efficiency remain to be validated through future commercial deployment.