Huawei Unveils OceanStor M900 Context Memory Storage for Scalable KV Cache

At the Huawei Connect 2026 event in Shanghai on September 17, Huawei ICT BG CEO Yan Chaobin introduced Lingqu UnifiedBus as the central interconnection element for a new joint cluster and SuperPoD architecture designed for Agentic AI workloads.

Huawei explained that as clusters grow, efficiency often decreases due to card communication latency. Furthermore, training models with 10 trillion parameters and frequent data exchange between agents and models push the KV Cache and intermediate data far beyond the memory of a single card, making the factory design a performance bottleneck.

The core principles of Lingqu's design include protocol unification: over ten interconnect protocols are merged into a single factory with Lingqu memory semantics. Interconnect bandwidth increases from the 100 GB class to the TB class, and round-trip time is reduced from approximately 7 microseconds to 2 microseconds, enabling global memory access within the SuperPoD.

Central Processing Units (CPUs), Neural Processing Units (NPUs), memory, and Solid State Drives (SSDs) are connected as equal elements for decentralized access. A flexible ratio of CPU to NPU is provided, along with hardware acceleration for Attention and FFN splitting for AF-distributed deployment. Multi-level storage pools can handle activations and use DDR as secondary memory for NPUs, while the optical network is positioned as a high-speed, low-latency data transmission channel for elastic scaling and expansion.

Huawei also announced multi-level Lingqu interconnect hardware, covering rack, inter-rack, and cluster levels. Modules within the rack eliminate losses from copper cables and circuitry—Huawei claims that a 4096-card SuperPoD can save about 196 kilometers of copper cable. Inter-rack switches provide 176 ports with 1.6 Tbps bandwidth each, delivering 280 TB of optical bandwidth per chassis at an RTT delay of about 2 microseconds. The Lingqu Xinghe UBG network switches advertise a branching factor of 1024, designed to support the construction of SuperClusters with millions of cards as models grow to tens of trillions of parameters.

At this factory, Huawei described the Agentic SuperPoD cluster, which combines Kunpeng 950, Ascend 960 SuperPoD, OceanStor M900 memory storage, and UBG switches for heterogeneous computing and unified resources. A similar interconnection story extends to devices: the Atlas 650E air-cooled server can directly connect 16 NPUs across two nodes without a switch, functioning as a small SuperPoD for on-site trillion-parameter inference. The emphasis is placed on co-designing the system and interconnects, not just the near-packet optics of the Ascend 960 SuperPoD, and this remains a vendor roadmap statement until independent cluster measurements are available.

At the Huawei Connect 2026 event, held in Shanghai on September 17th, Chairman David Wang introduced the Ascend 960 SuperPoD as the next-generation supernode for artificial intelligence. The development focus is placed on interconnect scalability rather than single-chip performance.

According to the system description, the Ascend 960 SuperPoD is the first model to utilize Near-Packaged Optics (NPO). This technology integrates Huawei's Lingqu UnifiedBus factory with the Hi-ONE optical engine, enabling optical connections to be placed closer to the chip package.

According to First Finance and related reports, one Ascend 960 SuperPoD can connect approximately 4096 cards with a round-trip latency approaching 2 microseconds. Huawei asserts that these metrics allow for large-scale training and inference of models up to 10 trillion parameters.

Furthermore, broader delivery information was disclosed: Ascend SuperPoD systems have already been shipped to over 1000 customers from more than 370 companies, indicating the architecture's transition from demonstration stands to commercial use.

The chip roadmap has been adjusted. Huawei announced that the Ascend 960DT is planned for release in the first quarter of 2027, and the Ascend 960PR in the third quarter. The liquid Atlas 960 SuperPoD is also scheduled for release in the third quarter of 2027, three quarters earlier than the initial public forecast, which pointed to the end of 2027 for the Ascend 960. Wang also stated that Huawei has developed 11 UnifiedBus-based chips for large systems and confirmed the long-term goal of creating a million-card SuperCluster.

The main emphasis is on system engineering: this includes NPO optics, UnifiedBus, cooling systems, and cluster software, which allow multiple Ascend cards to function as a single machine. Reuters reports confirm the dual 2027 launch dates and the figures of over 1000 supernodes and more than 370 customers, but did not name the buyers. This news differs from Huawei's 'Intelligent World 2035' report published recently, as it presents specific information about the SuperPoD interconnect and the revised Atlas 960 schedule for operators who are already evaluating Ascend clusters.