#China #Shanghai #Huawei #AI Inference #OceanStor

Huawei's OceanStor M900: A Game Changer for AI Inference

At the recent HUAWEI CONNECT 2026 held in Shanghai, Huawei's Deputy Chairman, David Wang, unveiled the groundbreaking OceanStor M900 Context Memory Storage, a significant milestone aimed at revolutionizing AI inference within hyperscale data centers. This innovative product is designed to enhance the efficiency of AI computations by providing SuperPoDs (Super Processing-on-Demand) with seamless access to a fully shared memory space, significantly affecting the industry's approach to AI infrastructure.

Transitioning the AI Landscape

The past year has marked a rapid transition in AI technologies, moving from simple applications like chatbots to advanced, autonomous agents that can execute complex tasks across various sectors. This ongoing evolution signifies the dawn of what has been termed the "agentic AI era," where increasingly sophisticated AI applications are being integrated into critical business processes.

As large AI models surpass a staggering trillions of parameters, the need for robust infrastructure becomes evident, establishing SuperPoDs as the optimal platform for AI tasks. These requirements have exceeded traditional on-chip memory and DRAM capabilities, pushing their limits on capacity and affordability.

Pioneering Features of OceanStor M900

The OceanStor M900 Context Memory Storage addresses these challenges with its advanced design and capabilities. Here are its three standout features:

1. Breaking Capacity Boundaries

The OceanStor M900 employs a high-speed UnifiedBus network, allowing global pooling and sharing of its K-V (Key-Value) cache. This new architecture expands the K-V cache from on-chip memory and DRAM to SSDs, enabling a single cluster to provide an astonishing 64 PB of capacity. As a result, NPU (Neural Processing Unit) resources can leverage terabytes of available K-V cache, dramatically improving processing efficiency and hit ratios.

2. Enhanced Inference Performance

This storage architecture is the first in the industry to integrate the CPU, network controller, and NAND controller units. By offering native K-V semantics, it facilitates a direct one-hop connection from the SuperPoD's NPU to SSDs, thereby eliminating latency caused by protocol conversions. This design achieves a remarkable reduction in access time—from milliseconds to just 60 microseconds—enhancing overall performance and productivity.

3. Cost-Effective AI Adoption

One of the key innovations of OceanStor M900 is its K-V aware adaptive storage technology. It intelligently predicts the lifecycles of cached data and redistributes it across different storage media. This innovative approach not only significantly extends SSD endurance but also decreases long-term costs associated with media replacements and operational management, thereby opening doors for more extensive AI deployment in various sectors.

Conclusion

As AI continues its integration into core business operations, the evolution of AI infrastructure is crystal clear. With a focus shifting towards better synergy between compute, network, and storage systems, solutions like the OceanStor M900 Context Memory Storage will play a critical role in enhancing data center efficiency. By enabling expansive capacity and streamlined access, Huawei is setting the stage for a new era in AI-driven technologies, making large-scale AI more accessible and economical than ever before.