#China #Shanghai #Huawei #AI Inference #OceanStor M900

Huawei's OceanStor M900: Revolutionizing AI Storage for Data Centers

On September 17, 2026, during HUAWEI CONNECT 2026, David Wang, the Vice Chairman and Rotating Chairman of Huawei, revealed the new OceanStor M900 Context Memory Storage system. Designed specifically for AI inference in hyper-scale data centers, this innovative product is anticipated to transform AI infrastructure by facilitating seamless collaboration between computing, network, and storage components.

A Shift in AI Infrastructure

As we witness an accelerated transition in AI technology, it has evolved from simple chatbots to sophisticated agents capable of performing complex tasks autonomously. These advancements have been widely adopted across critical sectors, signaling the onset of an agent-based AI era. With AI models reaching mammoth scales of up to 10 trillion parameters, the efficiency of data storage and retrieval has become increasingly significant, thus placing SuperPoDs at the forefront of AI infrastructure solutions.

SuperPoDs now support context windows exceeding one million tokens, a culmination of multi-turn inference and complex task execution. As the cache data generated during inference continues to rise, traditional memory systems including on-chip memory and DRAM are pushed to their limits in terms of capacity and performance. Recognizing this challenge, the industry has begun to converge on a solution involving multi-level storage systems capable of harmonizing memory types to create expansive shared memory environments.

Features of OceanStor M900

The OceanStor M900 is engineered to address the memory capacity bottlenecks that arise from ultra-long context needs and multi-turn inferences. Utilizing UnifiedBus networking, it establishes a global, multi-tier KV cache at petabyte scale, allowing for quick, single-hop connectivity. This architecture is designed to fully unleash the computing power of SuperPoDs, thus significantly enhancing AI inference capabilities in hyper-scale data centers.

1. Expanding Memory Capacity for Large-Scale AI

OceanStor M900 eliminates capacity limits by providing a massive shared memory solution. With UnifiedBus connectivity, the KV cache allows for global pooling and sharing through multi-tiered storage. This innovation enables a single cluster to reach a staggering storage capacity of 64 PB, significantly increasing the KV cache availability per Neural Processing Unit (NPU) from gigabytes to terabytes. This enhancement ensures better context storage, sharing, and reuse, thus improving cache hit rates dramatically.

2. Boosting Inference Performance to Maximize Computing Power

The OceanStor M900 boasts a groundbreaking architecture that combines the CPU, network controller, and NAND controller into a single entity, providing native KV semantics. This design facilitates a single-hop connection from the SuperPoD’s NPU directly to SSDs, eliminating protocol conversion requirements and minimizing access latencies from milliseconds to just 60 microseconds—a reduction of 90%. This efficiency leads to an aggregated access bandwidth of 40 TB/s, surpassing equivalent solutions by 1.5 times. In typical AI programming scenarios, this architecture doubles the token processing performance while halving the time to the first token, shifting compute power into actionable productivity more rapidly.

3. Reducing Token Costs for Wider AI Adoption

In a pioneering move, OceanStor M900 employs adaptive storage technology with KV recognition, intelligently predicting KV cache lifespan based on data value. This innovation allows for up to 24 drive writes per day, extending SSD longevity by 16 times and ensuring operational stability for three years. By reducing the frequency of media replacements and overall operational and maintenance costs, OceanStor M900 lowers the long-term expenses associated with large-scale AI inference frameworks, thereby fostering quicker transition to AI adoption across varied sectors.

The Future of AI Infrastructure

As AI systems increasingly integrate into production environments, we observe a significant shift from a computation-centric model to one focusing on deeper collaboration among computing, networking, and storage components. The role of context memory storage is poised to become essential in continually enhancing both capacity and efficiency for hyper-scale inference caches. With solutions like OceanStor M900, Huawei not only addresses existing challenges but also paves the way for future developments in AI technology and infrastructure.

In conclusion, Huawei's commitment to pioneering advancements in data center technologies is evident in the OceanStor M900, which stands as a testament to the ongoing evolution and growing importance of AI-driven systems across industries worldwide.