Huawei Repositions Storage as the Foundation of Enterprise AI
The idea that the real bottleneck of artificial intelligence infrastructures is no longer GPUs but rather how data is stored, retrieved, and protected is the thesis that Huawei brought to Paris on May 21, 2026, during the fifth Innovative Data Infrastructure Forum (IDI Forum). The event's slogan, "Data Awakening, Infra Evolving," summarizes the trajectory that the Chinese vendor intends to impose on its storage division, which aims to transform from a producer of storage systems to a supplier of specialized data infrastructures for AI.
The event, attended by Edge9 in Paris, brought together significant European clients, channel partners, analysts, and international press for one of the most articulated product launches in recent years within the category. Executives from Huawei along with researchers, end customers such as Auchan, Orange Business, and the National Library of Norway, and independent analysts, particularly Andy Buss, Senior Research Director at IDC, alternated on stage, framing the day within the ongoing transition from experimental generative AI to a phase of massive adoption based on agents.
The numbers cited by Buss and echoed by Huawei speakers leave little room for ambiguity about the scale of the phenomenon. Active AI agents in the world currently total about 30 million and are expected to surpass 2.2 billion by 2030; token consumption, an approximate indicator of inferential load, has jumped from about 6 to 15 billion per minute within a few months, and investment in AI by European companies is projected to quintuple from 2023 to 2029, approaching $100 billion. According to IDC, it is in this context that traditional IT architectures show their limits: high-quality data is scattered across silos, training pipelines cannot keep up with the pace imposed by distributed inference, and the lack of unified data governance results in an estimated 15% productivity loss for unprepared companies.
Five Pillars for the AI Data Center
On the technological front, during the keynote, Yuan Yuan, Vice President of Huawei and President of the Data Storage Product Line, presented the reorganization of the offer around five pillars, showcased as overlapping layers of a single stack for the AI data center: data lake, knowledge and memory platform, model engineering, agent framework, and data resilience. Each layer saw a timely announcement.
The first is AI Data Lake 2.0, a unified storage platform for consolidating training data and feeding models. The OceanStor Pacific Scale-Out system claims a density of 11 petabytes in 2 rack units with an average power consumption of 0.25 watts per terabyte, figures that Huawei claims are records in the category. Above the hardware level is DME Omni-Dataverse, the unified data space that the company says allows real-time visibility across up to 16 data centers, imports multimodal data, and performs vector searches on catalogs of over one hundred billion files in seconds. The promise, concretely developed in the case study of the National Library of Norway, presented by Marius Cusnes, head of the institute's IT platform, is to reduce latency between preservation archive and training environment. The library, custodian of about 20 petabytes of unique digital objects—stored across 60 petabytes due to replicated copies—has been tasked by the Norwegian government to create large national language models and now powers its AI pipelines with OceanStor Dorado all-flash systems.
The second pillar is the knowledge and memory platform, an area where Huawei places the most significant architectural innovations. The Context Memory Storage (CMS), introduced as the first solution of its kind designed for heterogeneous computing, offers shared memory at the petabyte level to third-party GPUs and Huawei's SuperPoD Atlas via direct semantic passage of key-value pairs. The goal, explained Yuan, is to reduce the TTFT, or time to first token generation, by an order of magnitude and cut the cost per token for large-scale inferential loads by 30%. For enterprise customers, where accuracy is prioritized over cost, Huawei instead offers the 3+1 AI Data Platform and the OceanStor A800 system, which combine KV cache acceleration, a knowledge base with declared retrieval accuracy above 95%, and a persistent memory bank for agents. The Unified Cache Manager (UCM) module, which orchestrates memory hierarchy between HBM, DRAM, and SSD, is presented as the lever that enables inferential accuracy to be increased by 30% compared to an architecture without semantic cache management.
Models, Agents, and Data Resilience
On the third pillar, dedicated to model engineering, Huawei proposes ModelEngine, a ready-to-use solution that integrates zero-code adaptation of new models, one-click release, and a partitioning mechanism for xPU with a ratio of up to 1:10. The idea is to allow a single accelerator to serve multiple loads simultaneously, especially in scenarios where models of different sizes, from a few billion parameters to the 671 billion in a clinical case cited on stage, must coexist on the same hardware pool. Above ModelEngine lies Nexent, the new framework for agents that, according to the company, generates agents from a description in natural language and reduces release times by 80%, evolving them autonomously thanks to automatic optimization of skills, prompts, and memory.
The fifth layer, data resilience, has a weight that the Paris forum has notably amplified. Benoit Heats, CTO of Huawei's Data Center Solutions division in France, detailed the company's new proposal within a seven-layer platform that integrates collaborative detection with firewalls, active protection via OceanCyber with a declared 99.99% accuracy against ransomware, an evolution of Air Gap 2.0 for clean room copying, and a clean recovery with isolated backup testing. The new generation of OceanProtect appliances, the X8100 and X9100 series, boasts a data reduction ratio of 90 to 1 and recovery speeds of one terabyte per second, aiming to make all-flash backup the default choice for businesses adopting AI workloads in production.
Virtualization, Data Sovereignty, and European Ambition
The presence of clients like Auchan and Orange Business was not just a backdrop. The French retail giant, detailed by Pierre-François Rougan, CTO of the group, has recently completed a migration of over 5,000 virtual machines from the previous hypervisor to Huawei's DCS platform, a full-stack virtualization solution that the Chinese vendor proposes as an alternative in the post-VMware phase. The project, based on a three data center architecture for disaster recovery, reported a 50% reduction in TCO over a five-year horizon.
Orange Business, through Christoph Lavan, head of cloud infrastructure for global delivery, outlined a consolidation of over 20,000 virtual machines and 10 petabytes of SAN flash storage on Dorado 8000 systems in active-active HyperMetro configuration across two sites, with an additional vertical offering for video surveillance based on the OceanStor 5500 family.
The question of how much priority Huawei intends to give to virtualization, in a market context marked by uncertainties regarding the Broadcom platform, was raised at the press round table. Jeff Wu, Chief Marketing Officer of the Data Storage Product Line, confirmed that DCS is now considered a strategic product, part of a full-stack vision that extends from primary storage to data lifecycle management through the management engine DME and the operation engine BME. The declared targets are cross-sector: production, finance and banking, public sector, and telecoms. In response to the objection that Huawei's European share in enterprise storage remains under pressure compared to Western vendors, Wu claimed the company's leadership position in the Gartner Magic Quadrant for enterprise storage platforms, the installed base of about 30,000 customers in 150 countries, and the presence of 53 of the top 100 banks worldwide in its portfolio.
More than on market share numbers, Huawei's European offensive relies on three levers. The first is the consolidation of the message "in Europe, for Europe," supported by Wei Song, President of Huawei European Enterprise Business, through the company's 29 research centers on the continent. The second is the alignment with the European AI plan, which foresees 19 national infrastructures dedicated to AI and five larger-scale centers under construction, creating a political space for private and sovereign architectures. The third lever, less emphasized but likely decisive, is the attempt to occupy the space of virtualization and resilience at a moment where many enterprise clients are re-evaluating the stack on which they have built the last decade. In the coming months, Huawei's ability to translate the ambition of Paris into a concrete share in European enterprise IT budgets will be measured on these three levers.