Huawei launched the OceanStor M900 context memory storage system on the 18th at HUAWEI CONNECT 2026 in Shanghai, designed for AI inference workloads in hyperscale data centers. The product provides fully shared petabyte-scale memory capacity for SuperPoD super-pod systems, with a single cluster delivering aggregate access bandwidth of 40 TB/s. Huawei Deputy Chairman and Rotating Chairman Wang Tao said in his keynote that 2026 marks an acceleration of AI from technological breakthroughs to large-scale deployment, with AI applications evolving from chatbots to intelligent agents and the arrival of the agentic AI era. As large language model parameter counts climb to the 10-trillion level, SuperPoD has emerged as the optimal choice for AI infrastructure, while mainstream large models commonly support context windows exceeding one million tokens. The OceanStor M900 employs Huawei's proprietary UnifiedBus high-speed interconnect, building a petabyte-scale global multi-tier KV cache through single-hop connections, enabling a single cluster to provide up to 64 PB of storage capacity. The KV cache capacity available to each neural processing unit jumps from gigabyte-level to terabyte-level. The system integrates the CPU, network controller module, and NAND controller module into a single framework, achieving single-hop direct connectivity from SuperPoD's NPUs to SSDs and compressing access latency from millisecond-level to 60 microseconds—a reduction of up to 90%.