Woofun AI reports that a stark divergence has emerged in the storage sector, with capital aggressively reevaluating AI-driven infrastructure while Web3 decentralized storage remains in a state of prolonged silence and decline. This disparity is anchored by the recent debut of Changxin, often cited as China’s first 'domestic storage stock,' on the ChiNext board, where it triggered explosive gains of 500%. Despite broader correction trends affecting the general storage sector, AI storage continues to command high valuations within the current technological narrative, whereas decentralized storage projects face long-term stagnation.

The root cause of this performance gap lies not in technical inferiority but in fundamental differences in value propositions: the market is currently rewarding 'hot data efficiency' while neglecting the 'trusted cold data' that underpins decentralized systems. Jacob Zhao of IOSG highlights this contrast, noting that while AI storage maximizes hash rate utilization and commercial profitability, decentralized storage adheres to principles of data fairness, censorship resistance, and the preservation of human civilization’s long-term memory.

The former serves as an efficiency-driven system for immediate computational needs, while the latter functions as a trustworthy foundation for immutable historical records. Although capital markets clearly favor the speed and scalability of AI infrastructure, the long-term necessity of an unalterable memory foundation remains intact, merely hidden in the shadows of current economic cycles.

The reevaluation of storage in the AI era represents a shift from traditional capacity-driven models to an efficiency-first paradigm. In the past, CIOs viewed storage as a cost center, focusing on unit costs, hard drive reliability, and replacement cycles spanning 3–5 years.

However, the rise of large models has transformed storage into a critical 'efficiency engine,' where the focus is now on optimizing GPU utilization, checkpoint writing speeds, and ultra-low latency for RAG applications. This evolution addresses the 'barrel effect' in AI infrastructure, where true hash rate utilization is determined by a complex multiplicative formula: True hash rate utilization = GPU × HBM × DRAM × SSD × network × file system. Any weakness in these components can lead to a collapse in overall performance, making storage a pivotal factor in commercial profitability.

In contrast, decentralized storage operates on a different value axis, prioritizing 'trusted cold data' that ensures data fairness and resists censorship. While AI storage is designed for the rapid processing of hot data, decentralized systems like Filecoin and Arweave are built to preserve data integrity over decades, serving as an unalterable foundation for human civilization. The current market neglect of decentralized storage does not diminish its long-term value but rather reflects a temporary preference for immediate computational gains over permanent data preservation.

The architecture of AI storage is a highly integrated, layered system that extends far beyond simple hardware components. Industry value and investor interest are concentrated on High Bandwidth Memory (HBM), enterprise-grade SSDs, SSD controllers, NVMe/CXL protocols, and high-performance storage systems. This architecture can be divided into four core layers, each serving a distinct function in the data flow.

The Compute-Proximate Memory Layer, or Bandwidth Core, relies primarily on HBM, supplemented by DRAM and CXL mempool technologies, to break through the 'memory wall' and ensure that hash rate potential is fully utilized. The High-Speed Persistent Storage Layer, or IO Hub, is built around enterprise-grade SSDs, which handle frequent checkpoint writes and large-scale training dataset loading.

The Low-Cost High-Capacity Storage Layer, or Capacity Foundation, utilizes HDDs and cold storage solutions to provide cost advantages for handling exponentially growing multimodal raw data and compliance logs. Finally, the AI Storage Systems and Data Software layer, or Scheduling Brain, includes parallel file systems and distributed object storage, transforming raw hardware into organized, indexed, and secured data services.

Decentralized storage does not compete in the millisecond-level speed races of these layers but instead establishes itself as a unique 'trusted cold layer,' focusing on data provenance and long-term archiving.

High Bandwidth Memory (HBM) serves as the 'bandwidth organ' closest to compute units in AI storage, playing a critical role in determining whether GPUs can receive sufficient data input. HBM is not traditional storage but a high-bandwidth memory layer located directly next to GPUs, designed to supply data at extremely high speeds. Its core architecture involves '3D DRAM stacking combined with 2.5D advanced packaging,' utilizing techniques like TSV vertical stacking and CoWoS heterogeneous integration to minimize the distance between storage and computing units.

This approach significantly boosts bandwidth but presents substantial barriers to entry, requiring expertise in DRAM manufacturing, TSV technology, ultra-thin stacking, packaging, heat dissipation, and customer certification. Currently, only three companies—SK Hynix, Samsung, and Micron—are capable of mass-producing HBM, giving them a competitive advantage in top-tier DRAM manufacturing and relationships with NVIDIA and AMD clients. The reliability of HBM is paramount, as any defect in the production stages can render an entire stack unusable, making it one of the most critical bottlenecks in the current AI supply chain.

This concentration of capability among a few players underscores the high stakes involved in the bandwidth core of AI storage architecture.

DRAM and Compute Express Link (CXL) form the system memory foundation and mempool engine for AI infrastructure. While HBM addresses extreme bandwidth requirements at the GPU level, DRAM serves as the basis for server system memory, handling CPU-side caching, data preprocessing, and temporary storage of intermediate results. The global DRAM market is dominated by SK Hynix, Samsung, and Micron, with Changxin (CXMT) playing a key role in China’s efforts to achieve domestic substitution for DRAM. CXL, or Compute Express Link, is a next-generation interconnect protocol designed to overcome the limitations of traditional DIMM slots and isolated server memory resources.

It promotes the evolution of memory architectures toward greater scalability, pooling, and sharing, allowing data centers to reshape how memory resources are organized. Although CXL is still in its early stages, transitioning from platform support to widespread deployment, it holds significant long-term architectural value. Key companies involved in this space include Astera Labs and Lanqi Technology, which are developing solutions to facilitate this transition. The integration of CXL with DRAM aims to break down physical limitations and create a more flexible and efficient memory ecosystem for AI workloads.

Enterprise-grade SSDs represent the high-throughput persistent backbone of AI data centers, composed of NAND chips, SSD controllers, and NVMe/PCIe data pathways. These components work together to deliver high throughput, low latency, and stable quality of service, enabling continuous data delivery to GPUs throughout the lifecycle of tasks such as training dataset loading, checkpoint writing, and RAG queries. The industrial chain for enterprise-grade SSDs is divided into three distinct stages: NAND chips, which determine storage density and cost; SSD controllers, which optimize performance and manage device lifespan; and NVMe/PCIe interfaces, which improve data transfer efficiency.

Notable companies in the NAND chip sector include Samsung, SK Hynix (through Solidigm), Micron, Kioxia, Western Digital, and YMTC. Leading SSD controller manufacturers include Phison, Silicon Motion, Marvell, and Maxio. In the transmission efficiency layer, Broadcom, Marvell, and Astera Labs are key players, particularly in the development of GPUDirect Storage technology, which reduces CPU involvement and alleviates I/O bottlenecks. This highly integrated system ensures that SSDs are not standalone hardware but critical components of the AI storage architecture, supporting the high-speed data flows required for modern AI applications.

Woofun AI data shows that the AI storage software stack acts as the control center for data availability, transforming underlying hardware into knowledge assets that can be directly utilized by AI systems. This stack consists of four main layers: High-Performance Storage Systems, Object Storage, Vector Databases, and the RAG Data Layer. High-Performance Storage Systems, provided by companies like VAST Data, WEKA, and Pure Storage, focus on high concurrency and low latency, addressing 'data hunger' in GPU clusters.

Object Storage, exemplified by AWS S3, centers around object, key, and metadata management, storing large volumes of unstructured data with an emphasis on cost efficiency and cloud-native features. Vector Databases, such as Pinecone and Milvus, are responsible for storing, indexing, and retrieving vectors generated by embedding models, enabling AI systems to accurately locate relevant information within vast knowledge bases.

The RAG Data Layer, represented by Databricks, goes beyond basic search functions by including data slicing, cleaning, access control, and provenance tracking, ensuring that enterprise data can be safely and traceably accessed by large models. This software layer is essential for organizing and securing data, making it accessible and usable for AI applications, and highlighting the importance of data management in the AI ecosystem.

Decentralized storage giants like Filecoin and Arweave offer alternative visions for data preservation, focusing on verification and permanence rather than speed. Filecoin has developed a comprehensive verifiable economic system through PoRep and PoSt mechanisms, positioning itself as a verifiable computing infrastructure. Instead of competing with AWS in consumer-grade cloud storage, Filecoin is shifting focus to hosting AI datasets and ensuring compliance-oriented archiving, providing verifiable trails for model auditing and copyright verification. The path forward for Filecoin involves developing S3-compatible APIs and supporting fiat currency payments, upgrading from a 'low-cost storage market' to a robust infrastructure for verifiable data.

Arweave, on the other hand, promotes the concept of 'one-time payment, permanent storage,' using mechanisms like Blockweave and SPoRA to incentivize miners to store and enable quick access to rare historical data. Its ideal role is as a foundation for human public memory, preserving records of human rights, war crimes, cultural artifacts, legal documents, and financial history. Arweave’s value lies not in speed but in its ability to preserve civilization’s memory across economic cycles, offering a permanent and censorship-resistant archive for humanity.

Despite their strong value propositions, decentralized storage projects face significant challenges, including mismatched supply and demand incentives, lack of enterprise-level services, poor search experience, insufficient privacy compliance, and token economy fluctuations. Early projects like Filecoin expanded rapidly using tokens but failed to create strong demand, resulting in massive capacity with low utilization and poor monetization.

These projects often reward those who can store data rather than those whose storage is truly needed, highlighting a gap between geeky ideals and mainstream commercial applications. The lack of enterprise-level services is another critical issue, as AWS’s competitive advantage lies in its 'data operating system,' which includes APIs, SLAs, access control, compliance auditing, and technical support. Enterprises seek reliability, not experimental infrastructure that requires them to manage keys and node selection on their own.

Additionally, the poor search experience in decentralized networks, due to distributed nodes and complex topologies, makes them unsuitable for handling AI hot data workflows. Privacy compliance is also a concern, as enterprise-sensitive data cannot be easily uploaded to public, permanent networks, creating a conflict between deletion rights and permanent immutability. Finally, the token economy amplifies cyclical fluctuations, with bull markets masking underlying demand issues and bear markets exposing weaknesses in commercial viability. Other decentralized storage projects, such as Storj, Sia, BNB Greenfield, Walrus, Celestia, EigenDA, and 0G, focus on specific ecosystems or niche areas, but their actual demand, developer adoption, and commercial viability remain to be proven.

The future of decentralized storage lies in balancing efficiency with immutable trust, as the balance of power will not always lean toward speed. Issues such as arbitrary content removal by tech giants, rising AI-related copyright disputes, concerns over data monopolies leading to loss of public archives, and increasing regulatory scrutiny of model training data could drive a renewed interest in decentralized solutions. As AI systems become more integrated into society, the need for a trustworthy, censorship-resistant foundation for data preservation will grow.

Decentralized storage, with its focus on data fairness and long-term memory, offers a unique value proposition that complements the efficiency-driven AI storage ecosystem. While current market mechanisms reward efficiency, the long-term necessity of an unalterable memory foundation remains intact. The hidden value of decentralized storage is not disappearing but waiting to be reassessed by future generations, as the importance of preserving human civilization’s memory becomes increasingly apparent in an era of rapid technological change. This marks a critical juncture where the industry must recognize the complementary roles of hot data efficiency and cold data trustworthiness in building a sustainable digital future.