During the Huawei Connect 2026 event, Chairman David Wang introduced the OceanStor M900 AI storage system, designed for SuperPoD clusters used for inference in hyperscale systems. This product aims to solve the 'memory wall' problem that arises when context windows reach one million tokens, and multi-iteration agent workloads increase the KV cache volume beyond the capabilities of HBM and DRAM.
Huawei positions the M900 as a petabyte-scale shared memory space, enabling AI infrastructure to transition from a compute-centric design to a co-design of compute resources, network, and storage. On the Lingqu OceanStor M900 platform, the KV cache is aggregated and distributed, allowing the SuperPoD memory to be expanded from device RAM to solid-state drives (SSDs).
The company claims that a single cluster can achieve a capacity of 64 petabytes, translating the available KV cache per NPU from gigabytes to terabytes and increasing the cache hit rate for long-context reuse. This approach views storage as a complement to inference memory, rather than a traditional backup or object layer.
Key performance claims are tied to the 'three-in-one' architecture, which includes CPU, network, and disk controller, with native support for KV semantics. Huawei states that NPUs can access SSDs in a single step without the need for protocol translation or routing through the CPU, reducing access latency from milliseconds to approximately 60 microseconds—a reduction of about 90%. Furthermore, the system provides an aggregate cluster throughput of about 40 TB/s, which is about 1.5 times higher than previous industry solutions. In a typical AI encoding scenario, Huawei asserts that token throughput can double, and time-to-first-token can be halved.
Cost control is achieved through KV-aware adaptive scheduling, which predicts the value and lifecycle of KV cache data across different media. Huawei claims the ability to achieve up to 24 DWPD and increase the lifespan of SSD media by approximately 16 times, supporting stable operation for three years with fewer drive replacements. Independent operators will still need to independently verify latency, throughput, and endurance metrics on their SuperPoDs, but the M900 is clearly positioned as memory infrastructure for agents and long context, complementing rather than replacing solutions based on Ascend chips or NPO interconnects.