Huawei Launches OceanStor M900 for Enhanced AI Performance in Large Data Centers
At the recent HUAWEI CONNECT 2026 event, David Wang, the Deputy Chairman of Huawei, introduced the groundbreaking OceanStor M900 Context Memory Storage. Designed specifically to address the demands of AI inference in hyperscale data centers, this innovative product enables SuperPods to achieve unprecedented performance with its expansive PB-scale memory capabilities.
This introduction marks a pivotal shift in AI infrastructure, transitioning from a traditional compute-centric approach to a more integrated model that emphasizes collaboration among compute resources, network, and storage systems. As we move deeper into 2026, AI has transitioned from merely a technological curiosity to a foundational element across various industries, showcasing its ability to perform complex tasks autonomously, which signals the dawn of the agentic AI age.
SuperPods have emerged as optimal AI infrastructure as AI models grow exponentially, with some reaching an astonishing 10 trillion parameters. As the complexity of AI tasks increases, typical on-chip memory and DRAM are struggling to keep pace, posing capacity and cost challenges. The consensus among industry experts is clear: there is a pressing need to develop a multi-tier storage system. Such a system would effectively coordinate on-chip memory, DRAM, and SSDs, creating a fully shared memory environment that efficiently meets growing demands.
The OceanStor M900 is Huawei's solution to overcome these challenges, specifically targeting the memory limitations associated with ultra-long context and multi-turn inference. By leveraging the UnifiedBus network, OceanStor M900 enables the establishment of a global, multi-tier KV cache that can expand memory capabilities to an impressive 64 PB in a single cluster. This enables each NPU to increase its KV cache capacity from gigabytes to terabytes, significantly enhancing the ability to store and retrieve contextual data.
One of the standout features of the M900 is its performance capabilities. It integrates CPU, network, and NAND controller units in an innovative architecture, allowing for seamless connectivity from SuperPod NPUs to SSDs without the need for protocol conversion. This drastic reduction in access latency from milliseconds to just 60 microseconds represents a remarkable 90% improvement. Consequently, clusters configured with OceanStor M900 can deliver an impressive 40 TB/s of access bandwidth, putting it substantially ahead of comparable solutions in the market. This architecture effectively doubles the throughput of token processing while halving the time to receive the first token, translating processing capabilities into tangible productivity gains.
Furthermore, the OceanStor M900 introduces the industry's first KV-aware adaptive storage technology. This cutting-edge feature predicts token lifecycle patterns based on their data value and utilizes smart distribution methods across different storage media. This leads to significant cost savings, with an astonishing capability of sustaining 24 drive writes per day, effectively boosting SSD longevity by a factor of 16 and guaranteeing operational stability for three years. This advancement dramatically reduces replacement costs, simplifying the financial landscape for large-scale AI infrastructure, making it economically viable and accelerating widespread AI adoption.
As AI enforcement into vital industrial systems becomes commonplace, the evolution of AI infrastructure will require a shift from a compute-centric model towards an integrated approach that tightly connects compute resources, networks, and storage systems. The emergence of context memory storage solutions like OceanStor M900 is set to redefine how inferring technologies work, ensuring continuous improvement in capacity and access efficiencies, which are indispensable for success in hyperscale data inference scenarios.
This introduction marks a pivotal shift in AI infrastructure, transitioning from a traditional compute-centric approach to a more integrated model that emphasizes collaboration among compute resources, network, and storage systems. As we move deeper into 2026, AI has transitioned from merely a technological curiosity to a foundational element across various industries, showcasing its ability to perform complex tasks autonomously, which signals the dawn of the agentic AI age.
SuperPods have emerged as optimal AI infrastructure as AI models grow exponentially, with some reaching an astonishing 10 trillion parameters. As the complexity of AI tasks increases, typical on-chip memory and DRAM are struggling to keep pace, posing capacity and cost challenges. The consensus among industry experts is clear: there is a pressing need to develop a multi-tier storage system. Such a system would effectively coordinate on-chip memory, DRAM, and SSDs, creating a fully shared memory environment that efficiently meets growing demands.
The OceanStor M900 is Huawei's solution to overcome these challenges, specifically targeting the memory limitations associated with ultra-long context and multi-turn inference. By leveraging the UnifiedBus network, OceanStor M900 enables the establishment of a global, multi-tier KV cache that can expand memory capabilities to an impressive 64 PB in a single cluster. This enables each NPU to increase its KV cache capacity from gigabytes to terabytes, significantly enhancing the ability to store and retrieve contextual data.
One of the standout features of the M900 is its performance capabilities. It integrates CPU, network, and NAND controller units in an innovative architecture, allowing for seamless connectivity from SuperPod NPUs to SSDs without the need for protocol conversion. This drastic reduction in access latency from milliseconds to just 60 microseconds represents a remarkable 90% improvement. Consequently, clusters configured with OceanStor M900 can deliver an impressive 40 TB/s of access bandwidth, putting it substantially ahead of comparable solutions in the market. This architecture effectively doubles the throughput of token processing while halving the time to receive the first token, translating processing capabilities into tangible productivity gains.
Furthermore, the OceanStor M900 introduces the industry's first KV-aware adaptive storage technology. This cutting-edge feature predicts token lifecycle patterns based on their data value and utilizes smart distribution methods across different storage media. This leads to significant cost savings, with an astonishing capability of sustaining 24 drive writes per day, effectively boosting SSD longevity by a factor of 16 and guaranteeing operational stability for three years. This advancement dramatically reduces replacement costs, simplifying the financial landscape for large-scale AI infrastructure, making it economically viable and accelerating widespread AI adoption.
As AI enforcement into vital industrial systems becomes commonplace, the evolution of AI infrastructure will require a shift from a compute-centric model towards an integrated approach that tightly connects compute resources, networks, and storage systems. The emergence of context memory storage solutions like OceanStor M900 is set to redefine how inferring technologies work, ensuring continuous improvement in capacity and access efficiencies, which are indispensable for success in hyperscale data inference scenarios.
Topics Consumer Technology）
【About Using Articles】
You can freely use the title and article content by linking to the page where the article is posted.
※ Images cannot be used.
【About Links】
Links are free to use.