From Hot Storage to Cold Memory: Decentralized Storage in the AI Era

Author: Jacob Zhao @ IOSG

Today, Changxin Storage, known as the "first domestic storage stock," officially listed on the ChiNext and exploded with an astonishing 500% surge. Although the storage sector has recently been affected by overall market corrections, AI storage is still being crazily revalued by capital in the current wave of technological narratives. Meanwhile, decentralized storage in the Web3 space has fallen into a long period of silence and loss. Why does the market performance differ so drastically for two entities both labeled as "storage"? The fundamental answer lies in the complete divergence of underlying value functions.

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, anti-censorship, and the long-term memory of human civilization. The former is an efficiency system for hot data, while the latter is a trustworthy system for cold data. The current capital market undoubtedly firmly stands on the side of "efficiency," but human civilization ultimately still needs an immutable memory foundation. The long-term value of trustworthy cold storage has never disappeared; it has merely been 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." CIOs focused on unit capacity costs, hard drive reliability, disaster recovery plans, archiving strategies, and equipment update cycles lasting 3 to 5 years. Storage was seen as an accessory following server purchases.

This round of storage boom is not a traditional cycle recovery but a repricing of data mobility capabilities driven by AI. In the era of large models, the logic of storage has transformed from "capacity first" to "efficiency above all," rigorously pursuing extreme indicators such as GPU feeding rates, Checkpoint writes, and RAG ultra-low latency. This marks a leap in storage value from "the final resting place of data" to "the high-speed passage for data entering computation."

The evolution of resource bottlenecks in AI infrastructure is essentially a battle to fill the "barrel effect." The true utilization rate of computing power is not a linear addition of single assets but a stringent multiplicative effect: true utilization rate of computing power = GPU × HBM × DRAM × SSD × network × file system; any shortcoming in one link will lead to the collapse of overall computing power utilization. In the AI era, storage has transformed for the first time from a "cost center" to an "efficiency engine." This is the fundamental logic behind the repricing of storage.

AI Storage Architecture Overview: From HBM Bandwidth Organs to Data Lake Foundations

AI storage is by no means a mere stacking of single hardware but a complex system of tightly coupled, layered scheduling. In this system, industrial value and capital focus are highly concentrated on HBM, enterprise-grade SSDs, SSD controllers, NVMe/CXL protocols, and high-performance storage systems. To clearly dissect its value flow, we categorize the AI storage architecture into four core levels from top to bottom:

• Compute Proximal Memory Layer (Bandwidth Core): Dominated by HBM, supplemented by DRAM and CXL memory pooling technology. This layer directly interfaces with GPU/CPU packaging or buses, aiming to break the "memory wall," serving as the first checkpoint determining whether computing power can be fully unleashed.

• High-Speed Persistent Storage Layer (IO Hub): The core logic is enterprise-grade SSD = NAND chips + SSD controller + NVMe/PCIe data pathways. This layer undertakes high-frequency Checkpoint writes, massive training set loading, and RAG hot data caching, representing the clearest persistent storage increment for AI data centers.

• Low-Cost Large Capacity Storage Layer (Capacity Foundation): Composed of HDDs, cold storage, and data lake archiving systems. In the face of exponentially expanding multimodal raw data, historical logs, and compliance backups, this layer still provides an irreplaceable TCO (Total Cost of Ownership) advantage.

• AI Storage Systems and Data Software (Scheduling Brain): Including high-performance parallel file systems, distributed object storage, vector databases, and RAG data governance layers. What AI truly consumes is not bare hardware but the data availability that has been efficiently organized, indexed, and authorized by the software stack.

As an ecological extension, decentralized storage does not directly engage in the millisecond-level competition of AI hot data but anchors on public dataset certification, AI training data provenance, and long-term cold memory archiving, establishing its unique ecological niche as a "trustworthy 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 proximal to the GPU. Its core mission is not to store data but to continuously "feed" data to computing power at extremely high bandwidth. HBM is the closest link to computing power in the AI storage chain, with the highest certainty, directly determining whether the GPU can be "fed" adequately, making it the current core supply chain bottleneck.

The core architecture of HBM is "3D DRAM stacking + 2.5D advanced packaging": through TSV vertical stacking and CoWoS heterogeneous integration, it achieves an extreme compression of storage-computing distance, resulting in a generational leap in bandwidth. Its industrial barriers are not just DRAM design but also involve DRAM manufacturing processes, TSV, ultra-thin stacking, packaging, heat dissipation, testing, and customer certification as a system engineering. Any defect in one link can lead to the scrapping of the entire HBM stack.

Currently, only SK Hynix, Samsung, and Micron can stably mass-produce, establishing a triple moat of top-tier DRAM manufacturing processes, packaging capabilities, and NVIDIA/AMD customer certifications.

DRAM and CXL: System Memory Foundation and Memory Pooling Engine HBM addresses the extreme bandwidth needs of GPU proximity, DRAM solidifies the server system memory foundation, while CXL attempts to break physical boundaries and reconstruct the organization of memory resources in data centers.

• DRAM: Primarily supports 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; Changxin Storage (CXMT) is a core variable for domestic DRAM substitution in China.

• CXL (Compute Express Link): A next-generation cache coherence interconnect protocol for data centers, aimed at breaking the limitations of traditional DIMM slots, local memory capacity, and server memory resource islands, promoting the evolution of memory architecture towards expansion, pooling, and sharing. Currently, CXL is still in the early stages of transitioning from platform support to large-scale deployment, with high long-term architectural value; core companies include Astera Labs and Lanqi Technology.

Enterprise-grade SSD: The Data Hub Built by NAND, Controllers, and NVMe Enterprise-grade SSDs are the most critical high-throughput persistent increments in AI data centers, continuously "feeding" data to GPUs with extremely high throughput, low latency, and stable QoS, spanning the entire lifecycle of training data loading, Checkpoint writing, RAG retrieval, inference caching, and log backflow.

In the AI storage architecture, SSDs are not isolated hardware but a highly coupled system, which can be distilled into an industrial formula: enterprise-grade SSD = NAND chips + SSD controller + NVMe/PCIe data pathways. The three layers represent independent links in the industrial chain:

• NAND Chips (Raw Material Layer): Determine storage density and unit cost, while the controller manages performance release and lifespan. Representative companies include Samsung, SK Hynix (Solidigm), Micron, Kioxia, Western Digital, and Yangtze Memory Technologies.

• SSD Controller (Performance Empowerment Layer): Determines performance release, data error correction, QoS stability, and wear leveling. Representative companies include Phison, Silicon Motion, Marvell, and Maxio.

• NVMe/PCIe (Data Pathway Layer): Determines the transmission efficiency of data from storage to computation. Combined with GPUDirect Storage technology, it reduces CPU memory bounce buffer and CPU involvement, significantly alleviating I/O bottlenecks. Representative companies include Broadcom, Marvell, and Astera Labs.

HDD / Cold Storage / Archiving: The Low-Cost Foundation of AI Data Lakes AI will not eliminate HDDs. With the demand for video and image data from multimodal large models and the exponential expansion of enterprise compliance logs and historical datasets, the need for low-cost cold data storage is surging. In the AI storage architecture, SSDs and HDDs collaborate based on business value: SSDs handle hot data and high throughput, while HDDs are responsible for low cost and long-term preservation. Representative companies include Seagate, Western Digital, and Toshiba. AI Storage Software Stack: The Scheduling Hub of Data Availability What AI truly consumes is never bare disks but "data services" that have been meticulously organized by the software stack. This architecture transforms underlying hardware into knowledge assets that AI can directly call upon, divided into four layers:

• High-Performance Storage Systems (Supply Systems): Focused on concurrent throughput and low latency, solving the "data hunger" problem of GPU clusters through parallel file systems, ensuring rapid flow of training and inference. Representative companies include VAST Data, WEKA, and Pure Storage.

• Object Storage (Raw Data Lake): Centered on Object, Key, and Metadata management, carrying massive amounts of 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 Databases (Semantic Indexing Layer): Responsible for storing, indexing, and retrieving vectors generated by embedding models, allowing AI to precisely locate relevant content from vast knowledge. Representative companies include Pinecone and Milvus.

• RAG Data Layer (Knowledge Invocation Layer): Beyond single retrieval, it encompasses data slicing, cleaning, permission control, and citation tracing, ensuring that enterprise data can be safely, accurately, and traceably invoked by large models. Representing companies: Databricks

From AI Hot Storage to Decentralized Cold Memory: Maximizing Efficiency vs. Maximizing Trust

AI Storage is an extremely efficiency-driven system, with its value function focused on maximizing computational output. HBM bandwidth determines whether GPUs can be adequately fed, while SSD throughput dictates the read/write efficiency of datasets and checkpoints; low latency is crucial for 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.

In contrast, the value function of decentralized storage is entirely different. It questions whether data will still exist in ten years, whether it has been tampered with, and whether it can withstand 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 irremovable historical records of human civilization. The former serves efficiency, pursuing extreme speed; the latter serves trust, defending silent memories. The former determines how fast models run, while the latter determines whether memories will be deleted. Currently, market mechanisms reward the efficiency of productivity, placing AI storage at the forefront, while decentralized storage seems to be experiencing a valuation collapse and a narrative drain.

The Vision and Reality of Decentralized Storage

There are numerous decentralized storage projects, but based on industry mindset and ecological sedimentation, the core representatives remain Filecoin and Arweave. Although both belong to "decentralized storage," their underlying architectural philosophies are almost two completely different paths— the former approaches AWS's elasticity through market contracts, while the latter approaches the eternity of libraries through a one-time social contract.

• Filecoin: It has built the most complete verifiable economic system through PoRep and PoSt. It should not continue to compete with AWS on consumer-grade cloud storage but should shift towards AI data provenance, public dataset hosting, and compliant archiving, providing verifiable chains for model auditing and copyright proof. The necessary path is to encapsulate 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 "one-time payment, permanent storage," it incentivizes miners to save and quickly access as much, especially scarce, historical data as possible through Blockweave and SPoRA mechanisms. Its best position is as a foundation for human public memory—preserving human rights records, war crimes evidence, cultural classics, archiving legal and financial history, and providing AI Agents with permanently accessible long-term memory. The value of Arweave lies not in speed, but in its capacity to carry civilizational memory across cycles.

The dilemma faced by decentralized storage projects like Filecoin and Arweave does not stem from incorrect value propositions, but from long-term mismatches in productization, retrieval experience, real demand, and Token incentives. This reveals a significant gap from geek ideals to mainstream commercial applications:

• Supply-Demand Incentive Mismatch: Early networks represented by Filecoin rapidly expanded through Tokens but failed to build a sufficiently strong demand side, resulting in huge capacity but inadequate utilization and payment conversion. They rewarded "I can store" rather than "I need to store".

• Lack of Enterprise-Level Service Capability: AWS's barrier is not the hard drive, but the "data operating system" composed of APIs, SLAs, permission management, compliance auditing, and technical support. Enterprises purchase "peace of mind," not experimental infrastructure that requires them to handle keys and node selection themselves.

• Retrieval Experience Shortcomings: "Storing in" does not equal "stably and with low latency retrieving out." Node dispersion, 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: Private enterprise data cannot simply be written into a public permanent network; the right to delete and permanent immutability are inherently in conflict. Decentralized storage is more suitable for public data and long-term archives, rather than indiscriminately accommodating core private data.

• Token Economy Amplification Cycle: Bull market financialization obscures insufficient demand, while bear market miner ROI declines expose commercialization shortcomings. Tokens can cold-start supply but cannot automatically create demand and sustainable income.

Other decentralized storage projects tend to focus on specific ecosystems or niche tracks: Storj/Sia's cross-cycle industry mindset and Web3 narrative influence are weaker than Filecoin/Arweave; BNB Greenfield/Walrus binds to specific public chain ecosystems of BNB or SUI; Celestia/EigenDA belong to the data availability (DA) layer, serving Rollup transaction confirmations rather than long-term archiving; projects like 0G that mix AI/DA narratives attempt to integrate storage, data availability, computation, and AI agent settlement into a set of AI-native modular infrastructure, but their real demand, developer adoption, and commercialization closed loops still need verification.

Future Opportunities for Decentralized Storage: The Long-Term Pendulum of Efficiency and Trust

During the explosive period of technological dividends, capital frantically chases efficiency, with assets like GPUs and HBM being assigned extremely high premiums, while decentralized storage advocating "trust and fairness" is naturally marginalized. However, the pendulum of history will not remain forever on the efficiency side. Events such as unreasonable bans and content deletions by super platforms, the outbreak of AI copyright lawsuits forcing data source proofs, data sovereignty disputes triggered by geopolitical conflicts, the disappearance of public archives due to data monopolies, and regulatory pressures on the compliance of model training data may all brew a revaluation of "trusted storage," while the future opportunities for decentralized storage still have the chance to reflect unique value in the following directions:

• AI Data Provenance: Building "data lineage proof" in conjunction with cryptographic proofs to address regulatory and auditing pressures.

• Public Datasets and Civilizational Archives: Anchoring censored archives and cultural heritage, constructing irreplaceable and irremovable memories.

• Trusted Archiving and Compliance Evidence: Achieving trusted self-evidence through Hash evidence, providing high-level digital notarization.

• Integration of ZK/TEE/DID Technologies: Resolving privacy tensions, upgrading from a single "storage protocol" to a "trusted data infrastructure".

• Invisible Product Route: Providing S3-compatible APIs and fiat billing, allowing users to directly purchase "trusted archiving" services.

AI storage and decentralized storage—one pursues extreme efficiency, providing fuel for our journey into the future; the other defends silent memories, safeguarding our right to look back at the past. The current market rewards efficiency without reservation, making decentralized storage appear silent or even collapsing; however, as the AI era further amplifies data monopolies, copyright disputes, and the fragility of historical memory, decentralized storage may welcome a revaluation of value in the form of "trusted cold layers." Memories that cannot be easily erased by platforms, companies, or any single power may transform from romantic idealism and marginal beliefs into necessary infrastructure.