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

Author: Jacob Zhao @ IOSG

Today, "China's first domestic memory stock," Changxin Storage Technology (CXMT), officially debuted on the ChiNext board, igniting the market with a staggering 500% surge. Although the overall storage sector is still experiencing some residual volatility from recent corrections, AI storage continues to be frantically revalued by capital amid the current wave of tech narrative. At the same time, decentralized storage in the Web3 realm has fallen into a prolonged period of silence and despondency. Why do entities both bearing the name "storage" exhibit such a stark contrast in market performance? The fundamental answer lies in a complete divergence of their underlying value functions.

The revaluation of storage in the AI era is essentially a frenzy surrounding "hot data efficiency," serving the ultimate maximization of computing utilization and commercial realization. What decentralized storage adheres to, however, is 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; the latter is a trustworthiness system for cold data. The current capital market undoubtedly stands firmly on the side of "efficiency," but human civilization ultimately still needs a tamper-proof memory foundation. The long-term value of trustworthy cold storage has never disappeared; it merely lies dormant in the dark side of the cycle, waiting to be repriced by the times.

In the traditional IT era, storage was a "capacity business." Enterprise CIOs focused on cost per unit capacity, hard drive reliability, disaster recovery solutions, archival strategies, and equipment refresh cycles spanning 3-5 years. Storage was seen as an accessory accompanying server procurement.

This round of storage fervor is not a traditional cyclical recovery but rather a repricing of data flow capabilities by AI. In the era of large models, the storage logic has qualitatively shifted from "capacity first" to "efficiency supreme," relentlessly pursuing extreme metrics such as GPU feeding rate, Checkpoint write speeds, and ultra-low latency for RAG. This signifies that the value of storage is leaping from "the final parking place for data" to "the high-speed channel for data entry into computation."

The evolution of resource bottlenecks in AI infrastructure is essentially a battle to fill the gaps in the "barrel effect." The real utilization of computing power is not a linear sum of individual assets but a stringent multiplicative effect: Real Compute Utilization = GPU × HBM × DRAM × SSD × Network × File System. A shortfall in any single link can cause overall computing efficiency to collapse. 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 storage's repricing.

AI storage is by no means a mere pile of individual hardware components; it is a complex, tightly coupled, and hierarchically managed system. Within 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 deconstruct its value flow, we divide the AI storage architecture into four core hierarchical layers:

• Compute-Near Memory Layer (Bandwidth Core): Dominated by HBM, supplemented by DRAM and CXL memory pooling technology. This layer is directly attached to GPU/CPU packaging or the bus, aiming to break the "memory wall." It is the first gate determining whether computing power can be fully unleashed.

• High-Speed Persistent Storage Layer (I/O 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. It represents the most definitive increment in persistent storage for AI data centers.

• Low-Cost, High-Capacity Storage Layer (Capacity Base): Consisting of HDDs, cold storage, and data lake archival systems. Faced with the exponential growth of multimodal raw data, historical logs, and compliance backups, this layer still provides irreplaceable TCO (Total Cost of Ownership) advantages.

• AI Storage Systems & 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 the data availability efficiently organized, indexed, and secured by the software stack.

As an ecosystem extension, decentralized storage does not directly engage in the millisecond-level race of AI hot data. Instead, it anchors positions like public dataset attestation, AI training data provenance, and long-term cold memory archival, establishing its unique niche as the "trustworthy cold layer."

High Bandwidth Memory (HBM) is not traditional storage; it is a high-bandwidth memory layer proximate to the GPU. Its core mission is not to save data, but to continuously "feed" data to the compute units with extremely high bandwidth. HBM is the link in the AI storage chain that is closest to compute, with the highest certainty. It directly determines whether GPUs can be "fed adequately" and is the most critical supply chain bottleneck currently.

The core architecture of HBM is "3D DRAM stacking + 2.5D advanced packaging": Through TSV vertical stacking and CoWoS heterogeneous integration, the distance between memory and compute is compressed to the extreme, achieving generational leaps in bandwidth. Its industrial barriers encompass not just DRAM design, but a system engineering effort involving DRAM process technology, TSV, ultra-thin stacking, packaging, thermal management, testing, and customer certification. A defect in any single link can cause the entire HBM stack to be scrapped.

Currently, only three giants worldwide can stably mass-produce HBM: SK hynix, Samsung, and Micron, who have built a triple moat of top-tier DRAM process, packaging capabilities, and NVIDIA/AMD customer certifications.

HBM solves the extreme bandwidth problem near the GPU, DRAM solidifies the server's system memory foundation, while CXL attempts to break physical boundaries and reorganize memory resource allocation within data centers.

• DRAM: Primarily handles CPU-side caching, data preprocessing, intermediate state storage, and system operation. It is the most fundamental system memory layer in servers. The global DRAM market is highly concentrated among three giants: SK hynix, Samsung, and Micron; Changxin Storage (CXMT) is the core variable in China's domestic DRAM substitution strategy.

• CXL (Compute Express Link): A new-generation cache-coherent interconnect protocol for data centers, aiming to overcome the limitations of traditional DIMM slots, local memory capacity, and server memory resource silos, driving memory architecture towards expansion, pooling, and sharing. CXL is currently still in its early stages, transitioning from platform support to scaled deployment, but holds significant medium-to-long-term architectural value. Key companies include Astera Labs and Montage Technology.

Enterprise SSDs represent the most critical high-throughput persistent storage increment in AI data centers. They continuously "feed" data to GPUs with extremely high throughput, ultra-low latency, and stable QoS, covering the entire lifecycle including training data loading, Checkpoint writes, RAG retrieval, inference caching, and log streaming.

Within the AI storage architecture, SSDs are not isolated hardware but a highly coupled system, which can be distilled into an industrial formula: Enterprise SSD = NAND Flash + SSD Controller + NVMe/PCIe Data Path. These three layers represent independent industrial chain segments:

• NAND Flash (Raw Material Layer): Determines storage density and unit cost. Representative companies: Samsung, SK hynix (Solidigm), Micron, Kioxia, Western Digital, Yangtze Memory Technologies (YMTC).

• SSD Controller (Performance Enablement Layer): Determines performance delivery, data error correction, QoS stability, and wear-leveling. Representative companies: Phison, Silicon Motion, Marvell, Maxio (Rayson).

• NVMe/PCIe (Data Path Layer): Determines the efficiency of data transfer from storage to compute. Combined with technologies like GPUDirect Storage, it reduces CPU memory bounce buffers and CPU involvement, significantly alleviating I/O bottlenecks. Representative companies: Broadcom, Marvell, Astera Labs.

AI will not eliminate HDDs. With multimodal large models' voracious appetite for video and image data, and the exponential growth of enterprise compliance logs and historical datasets, demand for low-cost cold data storage is simultaneously exploding. In AI storage architecture, SSDs and HDDs collaborate based on a tiered business value approach: SSDs handle hot data and high throughput, while HDDs handle low cost and long-term preservation. Representative companies include Seagate, Western Digital, and Toshiba.

What AI truly consumes is never bare drives, but "data services" meticulously organized by the software stack. This architecture transforms underlying hardware into knowledge assets directly callable by upper-layer AI. It can be divided into four specific layers:

• High-Performance Storage Systems (Feeding System): Focus on concurrent throughput and low latency. Parallel file systems address the GPU cluster's "data hunger," ensuring rapid flow for training and inference. Representative companies: VAST Data, WEKA, Pure Storage.

• Object Storage (Raw Data Lake): Core concepts are Object, Key, and Metadata management, hosting massive unstructured data. It doesn't pursue extreme low latency but builds capacity bases with low cost and cloud-native characteristics. Representative company: AWS S3.

• Vector Databases (Semantic Indexing Layer): Vector databases store, index, and retrieve vectors generated by embedding models, enabling AI to precisely locate relevant content from vast knowledge bases. Representative companies: Pinecone, Milvus.

• RAG Data Layer (Knowledge Invocation Layer): Goes beyond simple retrieval, encompassing data chunking, cleaning, access control, and citation provenance, ensuring enterprise data can be safely, accurately, and traceably invoked by large models. Representative company: Databricks.

AI Storage is an extreme efficiency-driven system whose value function focuses on maximizing computational output. HBM bandwidth determines whether GPUs can be fed adequately, SSD throughput determines dataset and Checkpoint read/write efficiency, and low latency is critical for RAG and real-time inference experiences. These metrics ultimately converge into GPU utilization and per-Token cost, directly deciding the commercial profitability of AI applications. The ultimate goal of AI storage is not preservation but acceleration, serving productivity.

Conversely, the value function of decentralized storage is fundamentally different. It questions 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 openly accessible, permanently preserved public data foundation. Its ultimate goal is to defend the absolute truth and sovereign independence of data, serving fairness, censorship resistance, and civilizational memory.

AI storage is "hot storage" providing fuel for future productivity; decentralized storage is "cold memory" preserving undeletable historical records for human civilization. The former serves efficiency, pursuing extreme speed; the latter serves trustworthiness, defending silent memory. The former determines how fast models run; the latter determines whether memories might be deleted. Currently, market mechanisms reward productive efficiency. AI storage is at the forefront of the wave, while decentralized storage seems to be experiencing a silence of valuation collapse and narrative bloodletting.

There are numerous decentralized storage projects, but in terms of industry mindshare and ecosystem depth, the core representatives remain Filecoin and Arweave. Although both fall under "decentralized storage," their underlying architectural philosophies are almost entirely divergent paths—the former approximates AWS-like elasticity through market-based contracts, while the latter approximates the permanence of a library through a one-time social contract.

• Filecoin: Has built the most complete verifiable economic system through PoRep and PoSt. It should not continue to compete directly with AWS on consumer-grade cloud storage. Instead, it should pivot towards AI data provenance, public dataset hosting, and compliance archiving, providing verifiable chains for model auditing and copyright proof. The necessary path is to package itself as an S3-compatible API supporting fiat payments, upgrading from a "cheap storage marketplace" to "verifiable computational infrastructure."

• Arweave: With its "pay once, store forever" narrative, it uses Blockweave and SPoRA mechanisms to incentivize miners to preserve and quickly access as much data as possible, especially scarce historical data. Its best fit is as a public memory foundation for human civilization—preserving human rights records, evidence of war crimes, cultural archives, archiving legal and financial history, and providing permanent, accessible long-term memory for AI Agents. Arweave's value lies not in speed but in its capacity to carry civilizational memory across cycles.

The predicament of decentralized storage projects like Filecoin and Arweave lies not in a flawed value proposition, but in the long-term misalignment among productization, retrieval experience, real-world demand, and token incentives. This reveals a significant gap between geek idealism and mainstream commercial adoption:

• Supply-Demand Incentive Misalignment: Early networks like Filecoin rapidly expanded capacity via token incentives but failed to build sufficiently strong paid demand, resulting in vast capacity but low utilization and conversion. It rewards "I can store" rather than "need me to store."

• Lack of Enterprise-Grade Service Capabilities: AWS's moat isn't hard drives, but the "data operating system" composed of APIs, SLAs, access management, compliance auditing, and technical support. Enterprises buy "peace of mind," not experimental infrastructure requiring them to handle keys and node selection themselves.

• Retrieval Experience Shortcomings: "Storing in" does not equal "retrieving stably and with low latency." Distributed nodes, complex topology, and a lack of unified SLAs make it difficult to handle AI hot data workflows. It's more suitable for trustworthy cold archiving and data provenance.

• Inadequate Privacy & Compliance: Enterprise private data cannot simply be written to a public, permanent network; the right to be forgotten inherently conflicts with permanent immutability. Decentralized storage is better suited for public data and long-term archives, not indiscriminately hosting core private data.

• Token Economics Amplifying Cycles: Bull market financialization masks insufficient demand; bear market ROI decline for miners exposes commercialization shortcomings. Tokens can bootstrap supply but cannot automatically create demand and sustainable revenue.

Other decentralized storage projects mostly focus on specific ecosystems or niche segments: Storj/Sia have weaker cross-cycle industry mindshare and Web3 narrative influence compared to 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 archival; AI/DA hybrid narrative projects like 0G attempt to integrate storage, data availability, computation, and AI agent settlement into an AI-native modular infrastructure, but their real-world demand, developer adoption, and commercial closure remain to be validated.

During periods of technological红利爆发, capital frantically chases efficiency. Assets like GPUs and HBM are assigned extremely high premiums, while decentralized storage advocating "trust and fairness" is naturally marginalized. However, history's pendulum will not remain forever on the side of efficiency. Unjustifiable platform bans and content takedowns, AI copyright lawsuits forcing data provenance, geopolitical conflicts sparking data sovereignty disputes, data monopolies leading to the disappearance of public archives, and regulatory pressure for compliance auditing of model training data—all such events could potentially brew a repricing of "trustworthy storage." Decentralized storage's future opportunities may still manifest unique value in the following directions:

• AI Data Provenance: Combining cryptographic proofs to build "data lineage proofs," addressing regulatory and audit pressures.

• Public Datasets & Civilizational Archives: Anchoring censored archives and cultural heritage, building irreplaceable, undeletable memory.

• Trustworthy Archival & Compliance Attestation: Achieving trustworthy self-attestation through hash storage, providing high-grade digital notarization.

• Integration with ZK/TEE/DID Technologies: Mitigating privacy tensions, upgrading from a single "storage protocol" to "trustworthy data infrastructure."

• Invisible Product Route: Offering S3-compatible APIs and fiat billing, allowing users to directly purchase "trustworthy archival" services.

AI storage and decentralized storage—one pursues ultimate efficiency, providing fuel for our sprint into the future; the other defends silent memory, safeguarding our right to look back at the past. The current market unhesitatingly rewards efficiency, which is why decentralized storage appears silent, even collapsing. However, when the AI era further amplifies data monopolies, copyright disputes, and the fragility of historical memory, decentralized storage may experience a revaluation as a "trustworthy cold layer." Those memories that cannot be easily erased by platforms, corporations, or any single authority might transition from an idealistic romance, from a fringe belief, into necessary infrastructure.