# Huawei’s full stack AI data platform and agent foundry > Source: > Published: 2026-05-27 13:05:00+00:00 AI/ML # Huawei’s full stack AI data platform and agent foundry [Huawei](https://www.blocksandfiles.com/flash/2026/05/22/norways-2-petabytes-of-huawei-flash-storage-and-llm-training/5244910) announced a full stack of AI model training and inference data storage and pipeline ingest/acceleration systems plus agent runtime frameworks and more at its ID 2026 forum in Paris. It did not mention the AI factory term but the AI Data Platform it announced has the depth and breadth of AI factory-type announcements from other major server systems and storage suppliers such as Dell and HPE. It has its OceanStor Pacific data lake storage with Omni-Dataverse ingest source connectors, all-flash OceanStor Dorado and A800 to store hot data for training and inference, with a Unified Cache Manager (UCM) and KV Caching facilities, accelerated by an OceanDisk product. This is extended with AI agent development and operations frameworks, and supported by OceanProtect backup and archive storage. Yuan Yuan, VP and President of the Huawei Data Storage Product Line, said: "The next chapter of AI is data. Committed to technological innovation in data storage, Huawei will accumulate the experience of industrial AI adoption, and work closely with the entire industry to help customers accelerate their journey into the intelligent era.” Huawei has a 5-layer platform for AI, starting from a data lake, and progressing through its AI data platform, computing with heterogeneous xPU adaptation, model to an agent framework layer. The company makes its own CPUs and GPUs (Ascend and Atlas accelerators) and cannot partner with Nvidia, or any other US GPU maker, because of the geo-political situation between the USA and China. As a result it has developed its own technologies, including software frameworks, to achieve the same kind of AI model and agent facilities as Nvidia’s AI factory schemes. Bulk AI data is held in all-flash OceanStor Pacific scale-out storage, capable of having 11 PB capacity in a 2 RU chassis. There are high-performance 9928 systems and high-capacity 9926 systems. Real-time data ingest from many sources is enabled with a DME Omni-Dataverse product, which has multi-modal, cross-site, global data visibility, a unified data space, and manageability features. It supports, Huawei says, retrieval from hundreds of billions of 1,000-dimension vectors in seconds. The Pacific uses Huawei’s palm-sized, [61.44 and 122 TB SSDs](https://www.blocksandfiles.com/flash/2026/05/21/huaweis-new-stacking-tech-for-high-capacity-ssds/5244276) and [hardware compression card](https://www.blocksandfiles.com/data-protection/2026/05/22/huaweis-up-to-901-compression-card/5245077). An AI Co-pilot provides natural language-based system management with a dedicated Fault Agent detecting anomalous situations and helping with recovery. Warm data needed for AI training and inference is stored in all-flash Dorado systems, with hot data going to OceanStor A800, a specialized high-performance distributed file/object storage system built for AI/ML workloads; training, inference, checkpoints, and large datasets with small files. The A800, with DPUs and GPU Direct-like functionality, has massive bandwidth for loading training sets, fast checkpointing/resumption, vector/tensor/RAG data handling, and integration with GPU/xPU clusters. This storage product area also has Huawei’s Data Engine node which handles AI-specific data processing on top of the underlying Dorado storage. It reads data from Dorado, either block, file, or object protocols, and then generates and stores vector embeddings, RAG (Retrieval-Augmented Generation) pipelines, knowledge bases, and memory structures needed for AI models. As we understand it, the Data Engine Node is Huawei's way of making general-purpose OceanStor Dorado arrays AI-ready by offloading intelligence, vectorization, and acceleration tasks. It complements the more tightly integrated OceanStor A800, which combines storage + data engine capabilities in one appliance for maximum performance in new AI builds. Huawei has a Unified Cache Manager (UCM) layer using Dorado + Data Engine Node or A800 appliance storage and accelerating data access, with a Knowledge Base, KV Cache store and Memory Bank logically placed between the UCM and the storage boxes. This is a CMS, a Context Memory System. The Knowledge Base stores stores massive amounts of multimodal data (text, images, video, etc.) converted into high-dimensional vectors, and is said to have over 95 percent retrieval accuracy. It supports petabyte-to-exabyte data levels with clustering and can store tens of billions of thousand-dimensional vectors (or larger in clusters). The [KV Cache](https://www.blocksandfiles.com/data-management/2022/04/10/kvcache/1594619?_gl=1*2o3hse*_ga*MzkxNDQyMTIwLjE3NzcwMzc0NTc.*_ga_NSDTXHMMN0*czE3Nzk4NzIyOTQkbzc1JGcxJHQxNzc5ODcyMzEyJGo2MCRsMCRoMA.. ) stores vectors and context data needed in current AI processing, and can scale to petabyte-size. It provides, in Nvidia CMS terms, a [G3.5 KV Cache memory tier](https://www.blocksandfiles.com/ai-ml/2026/03/30/nvidia-and-its-partners-kv-cache-extenders/5209284), and can support Huawei’s Atlas SuperPoD AI server os a “third-party GPU SuperPod.” The third item, the Memory Bank, provides multi-session or long-term memory storage and management for AI agents. It enables them to be coherent across multi-turn conversations and, Huawei says, capable of iterative self-improvement. It supports personalized behavior, backtracking, and multi-agent collaboration. Huawei’s AI product set also includes a Model Engine AI tool chain and framework with model engineering, gateway, xPU partitioning and intelligent scheduling. The partitioning achieves an up to 1:10 ratio of xPU partitioning to make one xPU serve multiple purposes. A Nexent agent framework helps industry sector customers develop and adopt AI agents. It directly generates agents via natural language–based interaction, simplifying development, and cuts rollout time by up to 80 percent according to Huawei. OceanProtect all-flash systems look after data backup and archive. Their QLC SSDs have an adaptive SLC section for actively-accessed data in recovery situations. Data resilience is augmented with an OceanCyber data security appliance with an AI-powered detection card and clean room recovery facilities. Yuan said: ”AI is unlocking new opportunities for the IT industry", and Huawei wants to take advantage of that. There are more than 30,000 OceanStor customers world-wide. They will all receive Huawei’s AI messages. This is serious storage outside the USA, and Huawei is the number 2 supplier of storage worldwide. All AI server and storage system suppliers should have it on their competitive radar screens.