Huawei OceanStor M900 AI Storage Boosts SSD Lifespan 16x

Huawei introduced the OceanStor M900 AI memory storage system at Huawei Connect 2026 on September 17, 2026. This launch addresses a critical bottleneck in large-scale AI training by managing the massive data throughput required for trillion-parameter models. The system is designed to accelerate both training and inference for Agentic super-node clusters, which are essential for handling complex, long-sequence AI workloads.

New system targets AI training bottlenecks with 16x SSD endurance

The core innovation of the OceanStor M900 is its multi-layer KV Cache architecture. It features a 3.5-layer PB-level KV cache that stores key-value pairs for AI models. The multi-layer KV Cache architecture enables the system to manage the substantial memory requirements of large language models with greater efficiency than prior systems.

Key Specifications

• KV Cache Architecture: 3.5-layer PB-level with one-hop direct connection
• SSD Endurance: 16x increase in read/write lifespan via hybrid media
• Cluster Scale: Up to 1,000,000 Ascend super-node cards
• Interconnect: Lingqu UnifiedBus unifies multiple protocols

To achieve this, Huawei implemented the Lingqu UnifiedBus protocol. This bus unifies multiple interconnect protocols into a single standard, which reduces the overhead associated with protocol conversion. It enables a 'one-hop direct connection' between the compute nodes and the KV cache, minimizing latency.

The storage subsystem also sees significant durability improvements. By using hybrid media and optimized retention algorithms, the M900 claims a 16x increase in SSD read/write lifespan. The 16-fold increase in SSD endurance reduces the frequency of drive replacements and lowers the total cost of ownership for data centers supporting intensive AI workloads.

Cluster scalability is another major focus of the M900 announcement. Huawei demonstrated a two-layer CLOS four-plane network that supports up to 512,000 cards. When combined with a multi-track topology, the system can scale to support up to 1,000,000 Ascend super-node cards in a single cluster.