Author: Jacob Zhao @ IOSG
Today, the "first stock of domestic storage" CXMT officially landed on the ChiNext board, igniting the market with a stunning 500% surge. Although the storage sector as a whole is still affected by the aftermath of the correction, AI storage is being frantically revalued by capital in the current wave of technology narratives. At the same time, decentralized storage in the Web3 field has fallen into long-term silence and disappointment. Why do both bear the name "storage," yet their market performances are as different as fire and ice? The root cause lies in the complete divergence of the underlying value function.
The revaluation of storage in the AI era is essentially a carnival about "hot data efficiency," serving the ultimate maximization of computing power utilization and commercial monetization; while decentralized storage adheres to the value proposition of "cold data trustworthiness," defending data fairness, censorship resistance, and the long-term memory of human civilization. The former is an efficiency system for hot data, and the latter is a trust system for cold data. The current capital market undoubtedly stands firmly on the side of "efficiency," but human civilization ultimately still needs an immutable memory foundation. The long-term value of trusted cold storage has never disappeared; it merely lies dormant in the dark side of the cycle, waiting to be repriced by the times.
Why storage has become the focus of the AI industry chain again
In the traditional IT era, storage was a "capacity business." Enterprise CIO focused on unit capacity cost, hard disk reliability, disaster recovery solutions, archiving strategies, and equipment replacement cycles of 3–5 years. Storage was seen as an accessory following server procurement.
This storage boom is not a traditional cyclical recovery, but a repricing of data flow capabilities by AI. In the era of large models, the storage logic has shifted from "capacity first" to "efficiency first," focusing on extreme metrics such as GPU feeding rate, checkpoint writing, and ultra-low latency for RAG. This marks the value of storage leaping from "the final parking lot of data" to "the high-speed channel for data entering computation."
The evolution of resource bottlenecks in AI infrastructure is essentially a battle to fill the "barrel effect." The true utilization of computing power is not a linear superposition of single assets, but a strict multiplier effect: true computing power utilization = GPU × HBM × DRAM × SSD × Network × File System. Any shortcoming in one link will cause the overall computing power utilization to collapse. In the AI era, storage has transformed from a "cost center" to an "efficiency engine" for the first time. This is the fundamental logic behind the repricing of storage.
AI Storage Architecture Panorama: From HBM Bandwidth Organ to Data Lake Foundation
AI storage is by no means a simple stack of hardware, but a tightly coupled, hierarchically scheduled complex system. In this system, industrial value and capital focus are highly concentrated on HBM, enterprise SSDs, SSD controllers, NVMe/CXL protocols, and high-performance storage systems. To clearly break down its value flow, we divide the AI storage architecture into four core layers from top to bottom:
Compute-Near Memory Layer (Bandwidth Core): Dominated by HBM, supplemented by DRAM and CXL memory pooling technology. This layer directly interfaces with GPU/CPU packages or buses, aiming to break the "memory wall" and is the first checkpoint determining whether computing power can be fully unleashed.
High-Speed Persistent Storage Layer (IO Hub): The core logic is Enterprise SSD = NAND Flash + SSD Controller + NVMe/PCIe Data Path. This layer handles high-frequency checkpoint writes, massive training dataset loading, and RAG hot data caching, representing the most significant persistent storage increment in AI data centers.
Low-Cost High-Capacity Storage Layer (Capacity Foundation): Composed of HDDs, cold storage, and data lake archiving systems. Facing exponentially growing multimodal raw data, historical logs, and compliance backups, this layer still offers irreplaceable TCO (Total Cost of Ownership) advantages.
AI Storage Systems and Data Software (Scheduling Brain): Includes high-performance parallel file systems, distributed object storage, vector databases, and RAG data governance layers. What AI truly consumes is not bare hardware, but data availability efficiently organized, indexed, and permissioned by the software stack.
As an extension of the ecosystem, decentralized storage does not directly engage in the millisecond-level race of AI hot data. Instead, it anchors on public dataset certification, AI training data provenance, and long-term cold memory archiving, establishing its unique ecological niche as a "trusted cold layer."
HBM: The "Bandwidth Organ" Closest to Computing Power in the AI Storage Chain
High Bandwidth Memory (HBM) is not traditional storage, but a high-bandwidth memory layer near the GPU. Its core mission is not to store data, but to continuously "feed" data to the computing power at extremely high bandwidth. HBM is the closest and most deterministic link to computing power in the AI storage chain, directly determining whether the GPU can be "fed" and is currently the most critical supply chain bottleneck.
The core architecture of HBM is "3D DRAM stacking + 2.5D advanced packaging": through TSV vertical stacking and CoWoS heterogeneous integration, the distance between storage and computing is extremely compressed, achieving a generational leap in bandwidth. Its industrial barrier is not just DRAM design, but a system engineering of DRAM process, TSV, ultra-thin stacking, packaging, heat dissipation, testing, and customer certification. Any yield defect in any link will cause the entire HBM stack to be scrapped.
Currently, only the three giants—SK Hynix, Samsung, and Micron—can achieve stable mass production, building a triple moat of top-tier DRAM process, packaging capabilities, and NVIDIA/AMD customer certification.
DRAM and CXL: System Memory Foundation and Memory Pooling Engine
HBM solves the near-GPU extreme bandwidth bottleneck, DRAM solidifies the server system memory foundation, and CXL attempts to break physical boundaries to restructure the organization of memory resources in data centers.
DRAM: Primarily handles CPU-side caching, data preprocessing, intermediate state storage, and system operation, serving as the most basic system memory layer for servers. The global DRAM market is highly concentrated among the three giants—SK hynix, Samsung, and Micron; CXMT (ChangXin Memory Technologies) is the core variable for China's DRAM domestic substitution.
CXL (Compute Express Link): A next-generation cache-coherent interconnect protocol for data centers, aiming to break through the limitations of traditional DIMM slots, local memory capacity, and server memory resource silos, driving memory architecture toward expansion, pooling, and sharing. Currently, CXL is still in the early stage of transitioning from platform support to large-scale deployment, with medium-to-long-term architectural value; key companies include Astera Labs and Montage Technology.
Enterprise SSDs: Data Hub Built by NAND, Controller, and NVMe
Enterprise SSDs are the most critical high-throughput persistent storage increment in AI data centers, continuously "feeding" data to GPUs with extremely high throughput, ultra-low latency, and stable QoS, spanning the entire lifecycle of training data loading, checkpoint writing, RAG retrieval, inference caching, and log feedback.
In the AI storage architecture, SSDs are not isolated hardware but a highly coupled system, which can be distilled into an industry formula: Enterprise SSD = NAND Flash + SSD Controller + NVMe/PCIe Data Path. The three layers represent independent industry chain segments:
NAND Flash (Raw Material Layer): Determines storage density and unit cost; the controller governs performance release and lifespan management. Representative companies: Samsung, SK hynix (Solidigm), Micron, Kioxia, Western Digital, YMTC.
SSD Controller (Performance Enabler Layer): Determines performance release, data error correction, QoS stability, and wear leveling. Representative companies: Phison, Silicon Motion, Marvell, Maxio.
NVMe/PCIe (Data Path Layer): Determines data transfer efficiency from storage to computation. Combined with GPUDirect Storage technology, it reduces CPU memory bounce buffers and CPU involvement, significantly alleviating I/O bottlenecks. Representative companies: Broadcom, Marvell, Astera Labs.
HDD / Cold Storage / Archiving: Low-Cost Foundation for AI Data Lakes
AI will not eliminate HDDs. With the insatiable demand of multimodal large models for video and image data, and the exponential growth of enterprise compliance logs and historical datasets, the need for low-cost cold data storage is surging. In AI storage architectures, SSDs and HDDs collaborate based on business value tiers: SSDs handle hot data and high throughput, while HDDs provide low cost and long-term retention. Representative companies include Seagate, Western Digital, and Toshiba.
AI Storage Software Stack: The Scheduling Hub for Data Usability
What AI truly consumes is never raw disks, but "data services" meticulously organized by the software stack. This architecture transforms underlying hardware into knowledge assets directly callable by upper-layer AI, specifically divided into four layers:
High-Performance Storage System (Feeding System): Focused on concurrent throughput and low latency, it solves the "data starvation" problem of GPU clusters through parallel file systems, ensuring rapid flow for training and inference. Representative companies: VAST Data, WEKA, Pure Storage.
Object Storage (Raw Data Lake): Centered on Object, Key, and Metadata management, it hosts massive unstructured data. It does not pursue extreme low latency but builds a capacity foundation with low cost and cloud-native characteristics. Representative company: AWS S3.
Vector Database (Semantic Index Layer): Vector databases store, index, and retrieve vectors generated by embedding models, enabling AI to precisely locate relevant content from vast knowledge. Representative companies: Pinecone, Milvus.
RAG Data Layer (Knowledge Retrieval Layer): Goes beyond single retrieval, encompassing data slicing, cleaning, permission control, and citation tracing, ensuring enterprise data can be safely, accurately, and traceably invoked by large models. Representative company: Databricks.
From AI Hot Storage to Decentralized Cold Memory: Maximizing Efficiency vs. Maximizing Trust
AI storage is an extreme efficiency-driven system, whose value function focuses on maximizing computational output. HBM bandwidth determines whether GPUs can be fed, SSD throughput determines the read/write efficiency of datasets and checkpoints, and low latency affects the real-time experience of RAG and inference. These metrics ultimately converge into GPU utilization and unit Token cost, directly determining the commercial profitability of AI applications. The ultimate goal of AI storage is not preservation but acceleration, serving productivity.
The value function of decentralized storage is completely different. It asks whether data will still exist in ten years, whether it has been tampered with, and whether it can resist single-point censorship. Through cryptographic proofs and distributed networks, it builds an open-access and permanently preserved public data foundation. Its ultimate goal is to defend the absolute truth and sovereign independence of data, serving fairness, anti-censorship needs, and civilizational memory.
AI storage is the "hot storage" that fuels future productivity, while decentralized storage is the "cold memory" that preserves non-deletable historical records for human civilization. The former serves efficiency, pursuing ultimate speed; the latter serves trust, defending silent memory. The former determines how fast models run, the latter determines whether memory will be deleted. Currently, market mechanisms reward productivity efficiency, putting AI storage at the forefront, while decentralized storage seems to be experiencing valuation collapse and narrative bleeding.
Vision and Reality of Decentralized Storage
There are many decentralized storage projects, but in terms of industry mindshare and ecosystem accumulation, the core representatives have always been Filecoin and Arweave. Although both belong to "decentralized storage," their underlying architectural philosophies are almost two completely different paths—the former uses market-based contracts to approximate AWS's elasticity, while the latter uses one-time social contracts to approximate the permanence of a library.
Filecoin: Through PoRep and PoSt, it has built the most complete verifiable economic system. It should no longer compete head-on with AWS in consumer-grade cloud storage, but instead pivot to AI data provenance, public dataset hosting, and compliant archiving, providing verifiable chains for model auditing and copyright proof. The necessary path is to package it as an S3-compatible API and support fiat payments, upgrading from a "cheap storage market" to a "verifiable computing infrastructure."
Arweave: With the narrative of "pay once, store forever," it uses the Blockweave and SPoRA mechanisms to incentivize miners to preserve and quickly access as much data as possible, especially scarce historical data. Its best position is as the public memory base of human civilization—preserving human rights records, war crime evidence, cultural classics, archiving legal and financial history, and providing permanently accessible long-term memory for AI agents. Arweave's value lies not in speed, but in its ability to carry civilizational memory across cycles.
The dilemma of decentralized storage projects like Filecoin and Arweave is not that their value proposition is wrong, but rather the long-term mismatch between productization, retrieval experience, real demand, and token incentives. This reveals a huge gap from geek ideals to mainstream commercial applications:
Supply-Demand Incentive Mismatch: Early networks represented by Filecoin rapidly expanded capacity through tokens but failed to build sufficiently strong paid demand, resulting in huge capacity but low utilization and paid conversion. Rewards are for "I can store" rather than "I need to store."
Lack of Enterprise-Grade Service Capabilities: AWS's moat is not hard drives, but the "data operating system" composed of APIs, SLAs, permission management, compliance audits, and technical support. Enterprises buy "peace of mind," not experimental infrastructure that requires them to handle keys and node selection.
Retrieval Experience Shortcomings: "Storing in" does not equal "stable, low-latency retrieval." Decentralized nodes, complex topology, and lack of unified SLA make it difficult to support AI hot data workflows, making it more suitable for trusted cold archiving and data provenance.
Insufficient Privacy Compliance: Enterprise private data cannot simply be written into a public permanent network; the right to deletion inherently conflicts with permanent immutability. Decentralized storage is more suitable for public data and long-term archives, not for indiscriminately hosting core private data.
Token Economy Amplifies Cycles: Bull market financialization masks insufficient demand; bear market miner ROI decline exposes commercialization shortcomings. Tokens can cold-start supply but cannot automatically create demand and sustainable revenue.
Other decentralized storage projects mostly focus on specific ecosystems or niche tracks: Storj/Sia have weaker cross-cycle industry mindshare and Web3 narrative influence than Filecoin/Arweave; BNB Greenfield/Walrus are tied to specific public chain ecosystems like BNB or SUI; Celestia/EigenDA belong to the data availability (DA) layer, serving Rollup transaction confirmation rather than long-term archiving; hybrid AI/DA narrative projects like 0G attempt to integrate storage, data availability, computation, and AI agent settlement into an AI-native modular infrastructure, but their real demand, developer adoption, and commercial closed loop remain to be verified.
Future Opportunities for Decentralized Storage: The Long-Term Pendulum of Efficiency and Trust
During the technology dividend explosion period, capital frantically chases efficiency, with assets like GPU and HBM commanding extremely high premiums, naturally marginalizing decentralized storage that advocates "trust and fairness." However, the pendulum of history will not stay on the efficiency side forever. Events such as unreasonable bans and content deletions by super platforms, AI copyright lawsuits forcing data provenance proof, geopolitical conflicts triggering data sovereignty battles, data monopolies leading to the disappearance of public archives, and regulatory audit pressure on model training data compliance may all brew a repricing of "trusted storage." The future opportunities for decentralized storage still have the potential to demonstrate unique value in the following directions:
AI Data Provenance: Combine cryptographic proofs to build "data lineage proofs" to address regulatory and audit pressures.
Public Datasets and Civilization Archives: Anchor censored archives and cultural heritage to build irreplaceable, non-deletable memory.
Trusted Archiving and Compliance Evidence: Achieve trusted self-certification through hash evidence, providing high-level digital notarization.
Integration of ZK/TEE/DID Technologies: Resolve privacy tensions, upgrading from a single "storage protocol" to a "trusted data infrastructure."
Invisible Product Approach: Provide S3-compatible APIs and fiat billing, allowing users to directly purchase "trusted archiving" services.
AI storage and decentralized storage: one pursues extreme efficiency, fueling our rush to the future; the other defends silent memory, preserving our right to look back at the past. The current market unreservedly rewards efficiency, leaving decentralized storage quiet or even collapsing. But when the AI era further amplifies data monopolies, copyright disputes, and the fragility of historical memory, decentralized storage may see a value reassessment as a "trusted cold layer." Those memories that cannot be easily erased by platforms, companies, or any single power may transform from idealistic romance and fringe belief into essential infrastructure.
