智东西（公众号：zhidxcom）
作者 | 杨京丽
编辑 | 李水青
智东西10月30日报道，昨日，2026数据存储产业大会上，数据存储专委会发布《数据存储产业发展研究报告（2026年）》和《数据存储产业与技术全景图（2026版）》。
报告预计，截至2026年底，全球存储产业规模将突破1.7万亿美元（约合人民币11.3万亿元），存储产业的增长重点将从单纯扩充容量，转向企业级SSD、AI数据中心和面向推理的数据基础设施。
会上，华为公司副总裁、数据存储产品线总裁袁远认为，GPU规模并不等于有效算力，AI发展需要提升存储对算力的支撑能力，加强数据治理和复用，让有限算力释放更高产出，同时把产业数据转化为AI应用优势。
他说：“数据是AI的燃料，数据决定了模型的高度，同时也决定了推理的准确度”。在他看来，存储一方面要提升算力利用率，减少重复计算；另一方面，要把原始数据清洗、整理成可直接调用的语料、知识和记忆，并真正用于模型训练、推理和Agent运行。
一、全球存储产业规模预计突破11.3万亿元，存储加速转向AI基础设施
《数据存储产业发展研究报告》围绕AI时代存储产业的发展趋势与技术路径展开分析。会上，中国电子工业标准化技术协会数据存储专业委员会秘书长孙钢对报告核心内容进行解读。
报告预计，截至2026年底，全球存储产业规模将突破1.7万亿美元（约合人民币11.3万亿元）。随着AI推理和智能体应用加速落地，存储产业的增长重点正从单纯扩充容量，转向企业级SSD、AI数据中心以及服务模型推理的数据基础设施。
报告提出，AI正在推动存储系统从单纯保存数据，转向同时管理数据、知识和记忆。在产业模式上，报告判断，存储正在从“建湖”转向“用湖”。过去，企业更关注把数据集中存起来；接下来，数据能否被检索、分析和复用，并真正服务于业务，将成为更重要的衡量标准。与此同时，存储交付将从一次性采购设备逐步转向按需提供服务，数据保护也要从单纯备份转向保障系统故障后快速恢复，尽量不影响业务运行。
二、70%算力尚未形成有效价值，存储需要支撑算力提效和数据治理
袁远谈道，当前国智算中心GPU平均利用率约为30%，约70%的算力尚未形成有效价值；企业真正能够使用的数据资源开发率约为28.4%，中文语料占全球比例约为4.8%。
在袁远看来，存储首先要解决数据供给、低成本留存和复用共享问题，减少重复计算和无效算力消耗。在推理环节，KV Cache需要低时延、高带宽和更低成本的存储支撑；在训练环节，随着模型参数规模继续扩大，Checkpoint效率和训练集群可靠性将成为关键瓶颈。
数据治理则对应语料、知识和记忆三类需求：数据要经过清洗、转换和标注，才能成为高质量语料；行业经验和实验数据需要进一步知识化；Agent运行过程中产生的上下文和经验，则需要通过存储形成可管理的长期记忆。
围绕AI应用的不同阶段，袁远提出“All in One”和“One for All”两条路径。“All in One”面向金融、制造、医疗等行业的AI使用端，以统一的AI数据平台承载知识沉淀、记忆管理、KV Cache加速和资源调度；“One for All”则面向模型厂商、互联网企业和国家级人工智能训练基地等AI生产端，建设面向超大规模训练和推理的专用数据基础设施。
结语：AI存储竞争转向数据可用性和系统效率
存储产业正在从提供容量，转向同时承担数据供给、算力协同和知识管理。对AI企业来说，存储系统能否让数据更快被调用、缓存更高效地复用，将直接影响训练和推理效率。
随着AI应用从模型训练扩展到推理和智能体运行，存储厂商的竞争重点也会从单一硬件性能，转向软硬件协同、数据治理能力和场景适配能力。谁能降低数据处理和算力调用的成本，谁就更有机会把AI能力真正落到行业应用中。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
YANG Kyung-Li
Editor, Lee Shui-qing
J.B., 30 October, yesterday, 2026 Data Storage Industry Congress, Data Storage Panel publish, Data Storage Industry Development Study (2026) and Data Storage Industry and Technology Panorama (2026 edition).
The report projects that by the end of 2026, the global storage industry will exceed $1.7 trillion (approximately RMB 11.3 trillion) and that its growth focus will shift from mere capacity expansion to enterprise-level SSD, AI data centres and reasoned data infrastructure.
At the meeting, Yuan Yuan, Vice-President of the company and Managing Director of the Data Storage Products Line, argued that the size of the GPU was not equal to effective computing, and AI development needed to enhance the capacity of storage to support computing, enhance data governance and reuse, and allow limited computing to release higher output while translating industrial data into AI application advantages.
He said: “The data are AI's fuel, which determines the height of the model and the accuracy of the reasoning”. In his view, storage, on the one hand, would have to improve the utilization of computing power and reduce double-counting; on the other hand, it would have to wash raw data, organize them into directly usable language, knowledge and memory, and actually be used for model training, reasoning and Agent operations.
The global storage industry is expected to exceed $11.3 trillion and accelerate the transition to AI infrastructure
The Data Storage Industry Development Study analyses trends and technology pathways in the AI-era storage industry. At that meeting, Sun Gang, Secretary-General of the Special Committee on Data Storage of the China Electronic Industry Institute, interpreted the core of the report.
The report projects that by the end of 2026, the global storage industry will have exceeded $1.7 trillion (approximately RMB 11.3 trillion). With the acceleration of AI reasoning and AI agent applications, the growth focus of the storage industry is shifting from a mere expansion of capacity to an enterprise-level SSD, the AI data centre and the data infrastructure of service model reasoning.
The report suggests that AI is moving the storage system from simply preserving data to simultaneously managing data, knowledge and memory. In terms of industry models, the report judges that storage is shifting from “lake-building” to “lake-use”. In the past, enterprises have focused more on centralizing data; then, the ability to retrieve, analyse and reuse data and to truly serve business will become a more important measure. At the same time, storage delivery will gradually shift from one-time acquisition of equipment to the provision of services on demand, and data protection will quickly recover from simple back-up to security system failure, with minimal impact on operations.
II. 70% of the calculus is not yet valid and storage needs to support calculus efficiency and data governance
Yuan goes on to say that the current average utilization rate of GPU is about 30 per cent and that about 70 per cent of the computing power is not yet valid; the real availability of data resources for enterprises is about 28.4 per cent and the global share of Chinese is about 4.8 per cent.
In the long term, storage will first and foremost address data availability, low-cost retention and reuse sharing, reducing double counting and inefficient computing consumption. In the reasoning chain, KV Cache needs low latency, high bandwidth and lower cost storage support; in the training chain, as model parameters continue to grow in size, Checkpoint efficiency and the reliability of training clusters will be key bottlenecks.
Data governance is based on three types of needs: Data need to be cleaned, converted and labelled in order to be of high quality; industry experience and experimental data need to be further knowledgeable; and context and experience generated during Agent's operation require the storage of managed long-term memory.
Around the different stages of AI application, Yuan proposed two paths, “All in One” and “One for All”. “All in One” is intended for the AI end of the financial, manufacturing, medical and other industries, to carry knowledge sedimentation, memory management, KV Cache acceleration and resource movement on a unified AI data platform, and “One for All” for the AI production end of model manufacturers, Internet enterprises and national artificial intelligence training bases to build a dedicated data infrastructure for mega-training and reasoning.
Conclusion: AI storage competition to data availability and system efficiency
The storage industry is shifting from providing capacity to simultaneously assuming data availability, computing synergies and knowledge management. For AI enterprises, the ability of the storage system to allow data to be accessed faster and cache to be reused more efficiently will have a direct impact on training and reasoning efficiency.
As AI applications expand from model training to reasoning and AI agent operations, the competitive focus of storage manufacturers will also shift from single hardware performance to software and hardware synergy, data governance and landscape adaptation. Those who can reduce the costs of data processing and leveraging will have a better chance of actually putting AI capacity into industry applications.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.