#China #Shanghai #Data Centers #AI Performance #Huawei OceanStor

Huawei Unveils OceanStor M900 for AI in Hyperscale Data Centers

At the recently held HUAWEI CONNECT 2026, David Wang, Vice Chairman of Huawei, introduced the OceanStor M900 Context Memory Storage, a groundbreaking system designed to accelerate AI operations within hyperscale data centers. This new product is set to transform the infrastructure of AI, shifting from a compute-centric model to one that fosters deeper collaboration between computing, networking, and storage. The OceanStor M900 offers a fully shared memory space and boasts capabilities that can handle petabyte-scale capacity and terabyte-per-second performance, marking a significant milestone in data processing.

The Shift to AI Agent Era

In 2026, the world witnessed a pivotal shift as AI technologies moved from theoretical advancements to large-scale implementations. AI applications have evolved from simplistic chatbots to sophisticated agents capable of autonomously executing complex tasks. These agents have found use in critical sectors, heralding the dawn of an era where AI-driven agents are fundamentally changing various industries.

With AI models growing to accommodate up to 10 trillion parameters, the SuperPoD (Super Performance Operating Data) infrastructure has emerged as the optimal choice for AI implementations. Notably, leading AI models now manage contextual windows exceeding one million tokens, making multi-round outputs and complex tasks common. Consequently, the limitations of built-in memory and Dynamic Random Access Memory (DRAM) have prompted the need for a multi-tiered storage system that coordinates these technologies to create a cohesive shared memory space.

OceanStor M900: Addressing Memory Bottlenecks

Huawei's OceanStor M900 Context Memory Storage has been developed specifically to alleviate memory capacity restrictions inherent in long-context and multi-round logical outputs. By utilizing the UnifiedBus network, OceanStor creates a global, multi-tiered Key-Value (KV) cache at petabyte scale with single-channel connections. This innovation unlocks the full computing potential of the SuperPoD, expediting AI output within large-scale data centers.

Key Features of OceanStor M900

1. Capacity Expansion for Large-scale AI: The OceanStor M900 allows massive memory utilization, supporting extensive AI operations through its high-speed UnifiedBus KV cache. It facilitates the seamless integration of built-in memory, DRAM, and SSDs, enabling a single cluster to offer up to 64PB of capacity, significantly increasing the available memory per neural processor from gigabytes to terabytes.

2. Enhanced Output Performance: As a pioneering architecture in the industry, OceanStor M900 integrates central processing units, network controller blocks, and NAND controller blocks. This design provides built-in KV semantics, allowing direct single-channel connections from SuperPoD neural processors to SSDs, thereby reducing access latency from milliseconds to as low as 60 microseconds—a reduction of 90%. Each cluster achieves a collective access throughput of 40TB/s, outperforming peer solutions by 1.5 times, thus doubling marker cluster throughput and reducing time to first marker in typical AI programming scenarios.

3. Cost-Efficiency for AI Deployment: The OceanStor M900 employs an industry-first adaptive storage technology that supports KV, adjusting data management based on the value of the information. This mechanism allows for up to 24 write operations per day (DWPD), extending SSD lifespans by 16 times while ensuring operational stability over three years. By lowering replacement and maintenance costs, the OceanStor m900 reduces long-term expenses associated with expansive AI deployment, promoting quicker implementation cycles.

As AI continues to permeate critical manufacturing systems across various sectors, the foundational architecture of AI infrastructure is transitioning from computation-centric models to a more integrated collaboration model that emphasizes the symbiotic relationship among computing, networks, and storage. The context memory storage solution is poised to play a crucial role in augmenting capacity and access efficiency in hyperscale AI deployments.