Huawei Launches AI Data Platform

The platform's core hardware, the OceanStor A800, is engineered for high-throughput AI workloads. In the MLPerf Storage v1.0 benchmark, a key industry test for AI hardware, the A800 demonstrated it could support the data throughput requirements of 255 simulated NVIDIA H100 GPUs with a single system, achieving a stable bandwidth of 679 GB/s. This level of performance is critical for reducing the time it takes to read and write model checkpoints, which can decrease GPU waiting times and improve overall training efficiency by more than 30%. A key software innovation in Huawei's platform is the Unified Cache Manager (UCM). This system intelligently allocates data across different memory tiers—such as high-bandwidth memory (HBM), DRAM, and SSDs—based on the latency requirements of the AI task. This software-defined approach is particularly noteworthy as it helps to mitigate hardware constraints, such as limited access to the latest HBM chips, by optimizing the use of available memory resources. For AI inference, the UCM can significantly reduce latency and increase throughput. Huawei claims the UCM can lower AI inference latency by up to 90% and boost system throughput by as much as 22 times. This is achieved by creating a hierarchical structure for the key-value (KV) cache, a crucial component in large language model inference, which speeds up response times. The company has announced plans to open-source the UCM, which could foster broader adoption and development within the AI community. In the context of fintech, the platform's high performance has direct applications in areas like real-time fraud detection. Such systems require the ability to process vast amounts of transaction data with ultra-low latency to identify and block fraudulent activities as they happen. The high I/O operations per second (IOPS) and bandwidth of the underlying storage are crucial for feeding data to complex AI models, like graph neural networks, which are used to uncover sophisticated fraud patterns. From a system design perspective, Huawei's storage-centric approach to its AI platform presents a different architectural philosophy compared to competitors like AWS SageMaker and Google Cloud's Vertex AI. While AWS and Google offer deeply integrated suites of software tools for the entire machine learning lifecycle, Huawei is leveraging its expertise in high-performance hardware to address the data bottleneck in AI workloads. This positions them to cater to enterprises that need to process massive datasets for training and inference. The AI infrastructure market is experiencing a massive influx of investment, indicating strong confidence in the growth of this sector. In 2025, AI infrastructure companies raised an unprecedented $84 billion in venture capital across just 10 mega-rounds. Overall, AI-related startups attracted nearly half of all global venture funding in 2025, with a total of $202.3 billion invested in the sector. This intense investment activity, particularly in foundational infrastructure, underscores the critical role that platforms like Huawei's will play in the expanding AI landscape.