At a critical juncture where AI is fully transitioning toward inference and continuous agent operation, China's AI storage industry has reached a milestone moment, moving from "disorderly competition" to "ordered innovation."
At the recently held 2026 Data Storage Industry Conference in Zhengzhou, the Data Storage Professional Committee of the China Electronics Standardization Association (hereinafter referred to as the "Data Storage Committee"), jointly with member units including Sugon and Huawei, officially released two major group standards for AI storage and announced the commencement of related national standard compilation work.
Image source: Sugon
Transitioning from "Scale Expansion" to "System Foundation Building"
Currently, artificial intelligence is undergoing a paradigm shift from training to inference, and from single interactions to continuous agent operation. Jin Lei, Division Chief of the Electronic Systems Division at the Department of Electronic Information under the Ministry of Industry and Information Technology (MIIT), pointed out that storage is entering an AI-driven supercycle, with its value being redefined. The MIIT's Department of Electronic Information has identified the storage industry as a core direction for cultivating new quality productive forces, systematically incorporating it into the "15th Five-Year Plan for the Development of the Electronic Information Manufacturing Industry."
The conference organizers emphasized that while the industry faces historic opportunities, it also encounters growing pains. This year's primary strategy of "foundation building" aims to transform advanced storage blueprints into tangible industrial capabilities through innovation-driven development and standard-led guidance.
Industry Urgently Needs a Unified "Yardstick" for Capabilities
Addressing the disorder and confusion caused by the past "proliferation of diverse approaches," Dr. Kang Yuzhuo, Product Manager of the Distributed Storage Product Department at Sugon, stated in a media interview that while technology routes among various vendors are evolving rapidly, there is significant chaos in product terminology and specifications. Although all are labeled "AI storage," some focus on bandwidth while others emphasize small files, with varying test conditions and scenario assumptions leading to potential bias during user selection. Particularly given the real-world demand for large-scale domestic deployment, massive construction of intelligent computing centers, and large model platforms, the industry urgently needs a unified consensus to avoid conflicting narratives and friction in supply-demand communication, project acceptance, and system integration.
The two group standards released at the conference—Storage Capability Requirements for Artificial Intelligence Applications and Technical Requirements for Large Model Inference Storage Systems—are designed to break this fragmented evaluation landscape. These standards not only unify terminology and capability definitions, allowing users and vendors to communicate using the same yardstick, but also construct for the first time a five-dimensional capability architecture covering "AI acceleration, scalability, compatibility, reliability, and security," ending the previous fragmentation of evaluation dimensions.
Four Core Metrics for Storage in the AI Era
Dr. Kang Yuzhuo emphasized that the large scale of AI business, multi-machine multi-card concurrency, and unique data flow characteristics dictate that the industry cannot directly apply traditional storage evaluation systems.
He pointed out that, guided by the new standards, evaluating AI storage systems should focus on four dimensions:
- Concurrent Performance: Traditional systems compete on peak performance for a single client, whereas large model training requires simultaneous access by dozens or hundreds of computing nodes. The new standards require products to benchmark against real-world business stability under multi-computing-node concurrency, emphasizing linear scaling from petabyte (PB) to exabyte (EB) levels to reduce training task interruptions and computing waste.
- Metadata Processing Capability: In AI scenarios, the volume of massive small files, images, text, and annotation files can easily exceed hundreds of millions. Traditional storage suffers from stuttering and increased latency when facing billions of metadata entries. The new standards designate metadata throughput, latency, and scalability as core assessment capabilities for AI storage.
- Dynamic Adaptation to Sub-scenarios: Abandoning the blind pursuit of "one architecture achieving the highest metrics for all." The standards emphasize adaptation to different AI sub-scenarios: training scenarios prioritize high aggregate bandwidth and concurrent stability; dataset preprocessing focuses on import/export efficiency; archiving prioritizes capacity and cost; while inference scenarios are singled out as a separate standard, primarily considering their low-latency business models.
- Business Continuity Assurance: Traditional systems mainly emphasize fault tolerance after hardware failures. However, large-scale training tasks have extremely long cycles, where minor storage jitter or IO latency spikes can cause entire task interruptions or rollbacks. The new standards comprehensively introduce a business perspective, focusing on latency jitter during long-term IO under high pressure, data verification integrity, and business recovery efficiency after task interruptions.
Defining Dedicated Components and Technical Specifications for Inference, Promoting Interface Convergence
Regarding the challenge of differing architectures and know-how among large model vendors, Dr. Kang Yuzhuo stated that storage resides at the bottom of the IT stack. From a data perspective, the underlying requirements of different large models ultimately converge. For example, regarding GPU direct access (accelerator direct pass-through), all vendors share the requirement that data bypasses CPU and memory transfers, enabling direct connection between VRAM and underlying storage to address the persistent issues of excessive data path copying and high computational overhead.
Based on this convergence of commonalities, the new standards provide clear capability requirements for accelerator direct-pass storage, zero-copy, multi-accelerator concurrency, and resource isolation at the product R&D level. Simultaneously, for inference scenarios, they specify capabilities and interface standards for KV Cache cross-node offloading and sharing, prefix caching, as well as accelerator direct-pass knowledge bases and memory banks. This will drive storage to evolve from simple file/object hosting to specialized systems capable of handling inference intermediate cache knowledge and memory data, effectively reducing VRAM pressure.
Elevating Group Standards to National Standards, Feeding Back into R&D Systems
Completing standard compilation is only the first step; driving implementation is key.
Dr. Kang Yuzhuo proposed that the next steps involve conducting industry dissemination and interpretation of the standards to bridge the cognitive gap between suppliers and users. Vendors must feed back the standard requirements into their product R&D systems, integrating capability baselines and testing requirements into product planning, iteration, and factory testing processes to achieve true "benchmark self-testing." Additionally, third-party institutions or joint industry testing platforms are needed to build executable testing tools and public testing environments. Given the rapid iteration of AI technology, a mechanism for fast standard updates following market trends will be established in the future.
Notably, these two group standards were conceived with national standards as the ultimate long-term goal from the outset. Participating organizations include storage system developers, storage medium manufacturers, research institutes, large model application users, and standard management bodies, ensuring high universality. Currently, the Data Storage Committee has partnered with the China Academy of Electronics Standardization to officially launch the compilation of one national standard related to storage system reference architectures, facilitating the transition from group to national standards and filling a key gap in the AI storage national standard system.