Huawei Unveils AI Data Platform with Pioneering "3+1" Architecture to Overcome Inference Bottlenecks

At the MWC 2026, Yuan Yuan, President of Huawei's Data Storage Product Line, officially launched the AI Data Platform during the product and solution release event. The platform introduces the groundbreaking "3+1" architecture, designed to address three major bottlenecks in AI inference: hallucinations, poor response experience, and memory loss.

Currently, the industry tends to "focus more on training and less on inference," but inference is the key to the practical application of AI. Huawei’s AI Data Platform addresses three core elements—knowledge, KV Cache, and memory—optimizing storage and enhancing the user experience through Unified Cache Manager (UCM) technology, which enables intelligent scheduling to improve inference performance.

Specifically, the platform achieves over 95% retrieval accuracy in its knowledge base by converting resources such as text, images, and videos into vast, fine-grained knowledge through multimodal lossless analysis and token-level encoding. The PB-level KV Cache can store massive historical data, significantly expanding the context window for single-session AI customer service conversations. It also reuses historical KV Cache in multi-round conversations, reducing first-token latency by 90%. The memory library offers context memory management, accurately extracting historical data and experiences and consolidating them into recallable memories, enabling the model to evolve continuously, becoming "smarter with each use." UCM technology further optimizes inference by managing and scheduling the knowledge base, KV Cache, and memory library through a three-layer caching architecture.

Regarding deployment modes, the platform provides two options: integrated and separate deployments. The integrated deployment uses OceanStor A800 as the foundation, combining all capabilities for both performance and scalability. The separate deployment utilizes a "data engine node + OceanStor Dorado" architecture, allowing additional data engine nodes to be added to existing systems, protecting historical investments and ensuring a smooth transition for business transformation.