IOSG | From Hot Storage to Cold Memory: Decentralized Storage Amidst the Storage Boom in the AI Era

Written by: Jacob Zhao, IOSG

Recently, "the first domestic storage stock" CXMT officially listed on ChiNext, igniting the market with a stunning 500% surge. Although the storage sector recently remained disturbed by the aftershocks of a pullback, AI storage is still being crazily revalued by capital in the current wave of tech narratives. At the same time, decentralized storage in the Web3 field has fallen into long-term silence and loss. Why, both bearing the name "storage," is the market performance so different? 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 compute utilization and commercial monetization; whereas 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, the latter is a trustworthiness system for cold data. The current capital market undoubtedly stands firmly on the "efficiency" side, but human civilization ultimately still needs an immutable memory base. The long-term value of trustworthy cold storage has never disappeared; it is just lurking in the dark side of the cycle, waiting to be repriced by the era.

Why Storage Has Become the Focus of the AI Industry Chain Again

In the traditional IT era, storage was a "capacity business." Enterprise CIOs focused on unit capacity cost, hard drive reliability, disaster recovery solutions, archiving strategies, and equipment update cycles of up to 3–5 years. Storage was seen as an accessory following server procurement.

This round of storage hype is not a traditional cycle recovery, but a repricing of data flow capability by AI. In the era of large models, storage logic has qualitatively changed from "capacity first" to "efficiency supreme," fighting hard on limit indicators such as GPU feed rate, Checkpoint writing, and extremely low RAG latency. This marks that storage value is leaping from "the final parking place for 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 "bucket effect." The real utilization rate of compute is not a linear superposition of single assets, but a strict multiplier effect: Real Compute Utilization = GPU × HBM × DRAM × SSD × Network × File System. A shortcoming in any link will cause the overall compute utilization rate to collapse. In the AI era, storage has for the first time changed from a "cost center" to an "efficiency engine." This is the fundamental logic behind storage being repriced.

AI Storage Architecture Panorama: From HBM Bandwidth Organs to Data Lake Base

AI storage is by no means a pile of single hardware, but a complex system of tight coupling and layered scheduling. In this system, industry value and capital focus are highly concentrated on HBM, enterprise SSDs, SSD controllers, NVMe/CXL protocols, and high-performance storage systems. To clearly dismantle its value flow, we divide the AI storage architecture from top to bottom into four core layers:

• Compute-Proximate Memory Layer (Bandwidth Core): With HBM as the absolute main force, supplemented by DRAM and CXL memory pooling technology. This layer directly fits GPU/CPU packaging or bus, aiming to break the "memory wall," and is the first gateway determining whether compute power can be fully released.
• High-Speed Persistent Storage Layer (IO Hub): The core logic is Enterprise SSD = NAND Die + SSD Controller + NVMe/PCIe Data Path. This layer undertakes high-frequency Checkpoint writing, massive training set loading, and RAG hot data caching, and is the clearest persistent storage increment in AI data centers.
• Low-Cost High-Capacity Storage Layer (Capacity Base): Composed of HDD, cold storage, and data lake archiving systems. Facing exponentially expanding multi-modal raw data, historical logs, and compliance backups, this layer still provides an irreplaceable TCO (Total Cost of Ownership) advantage.
• AI Storage System 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 data availability efficiently organized, indexed, and permissioned by the software stack.
• As an ecological extension, decentralized storage does not directly involve itself in the millisecond-level race of AI hot data, but instead anchors public dataset certification, AI training data provenance (Provenance), and long-term cold memory archiving, establishing its unique ecological niche as a "trustworthy cold layer."

HBM: The "Bandwidth Organ" Closest to Compute in the AI Storage Chain

High Bandwidth Memory (High Bandwidth Memory) is not traditional storage, but a high-bandwidth memory layer proximate to the GPU. Its core mission is not to save data, but to continuously "feed" data to compute with extremely high bandwidth. HBM is the link closest to compute and with the highest certainty in the AI storage chain, directly determining whether the GPU can be "fed full," and is 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 extremely compresses the storage-compute distance, achieving a generational leap in bandwidth. Its industry 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, Micron can stable mass produce globally, building a triple moat of top-level DRAM process, packaging capability, and NVIDIA/AMD customer certification.

DRAM and CXL: System Memory Base and Memory Pooling Engine

HBM solves the extreme bandwidth proximate to the GPU, DRAM consolidates the server system memory base, and CXL attempts to break physical boundaries and reconstruct the organization of data center memory resources.

• DRAM: Mainly bears CPU-side caching, data preprocessing, intermediate state temporary storage, and system operation, and is the most basic system memory layer of servers. The global DRAM market is highly concentrated in the three giants SK hynix, Samsung, Micron; CXMT is the core variable for China's DRAM domestic substitution.
• CXL (Compute Express Link): It is a new generation cache coherence interconnect protocol for data centers, aiming to break the limitations of traditional DIMM slots, local memory capacity, and server memory resource silos, promoting memory architecture evolution towards expansion, pooling, and sharing. Currently, CXL is still in the early stage from platform support to scale deployment, with high mid-to-long-term architectural value; core companies include Astera Labs and Montage Technology.

Enterprise SSD: Data Hub Built by NAND, Controller, and NVMe

Enterprise SSD is the core high-throughput persistent increment in AI data centers, continuously "feeding" data to the GPU with extremely high throughput, extremely low latency, and stable QoS, throughout the entire lifecycle of training data loading, Checkpoint writing, RAG retrieval, inference caching, and log feedback.

In the AI storage architecture, SSD is not isolated hardware, but a highly coupled system, which can be refined into an industry formula: Enterprise SSD = NAND Die + SSD Controller + NVMe/PCIe Data Path. The three layers represent independent industry chain links:

• NAND Die (Raw Material Layer): Determines storage density and unit cost, controller controls performance release and lifespan management    Representative companies Samsung, SK hynix (Solidigm), Micron, Kioxia, Western Digital, Yangtze Memory Technologies
• SSD Controller (Performance Enablement Layer): Determines performance release, error correction, QoS stability, and wear leveling. Representative companies Phison, Silicon Motion, Marvell, Maxio.
• NVMe/PCIe (Data Path Layer): Determines the transmission efficiency of data from storage to compute. Combined with GPUDirect Storage technology, it reduces CPU memory bounce buffer and CPU involvement, significantly alleviating I/O bottlenecks. Representative companies: Broadcom, Marvell, Astera Labs

HDD / Cold Storage / Archiving: Low-Cost Base of AI Data Lake

AI will not eliminate HDD. With the demand of multi-modal large models for video and image data, and the exponential expansion of enterprise compliance logs and historical datasets, low-cost cold data storage demand is surging simultaneously. In the AI storage architecture, SSD and HDD collaborate based on business value layering: SSD is responsible for hot data and high throughput, HDD is responsible for low cost and long-cycle preservation. Representative companies include Seagate, Western Digital, Toshiba.

AI Storage Software Stack: Scheduling Hub for Data Availability

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

• High-Performance Storage System (Supply System): With concurrent throughput and low latency as the core, it solves the "data hunger" problem of GPU clusters through parallel file systems, ensuring ultra-fast flow of training and inference. Representative companies: VAST Data, WEKA, Pure Storage.
• Object Storage (Raw Data Lake): With Object, Key, and Metadata management as the core, it bears massive unstructured data. It does not pursue extreme low latency, but builds a capacity base with low cost and cloud-native characteristics. Representative company: AWS S3
• Vector Database (Semantic Index Layer): Vector database is responsible for storing, indexing, and retrieving vectors generated by embedding models, allowing AI to precisely locate relevant content from massive knowledge. Representative companies: Pinecone, Milvus
• RAG Data Layer (Knowledge Call Layer): Beyond single retrieval, it covers data slicing, cleaning, permission control, and citation provenance, ensuring enterprise data can be safely, accurately, and traceably called by large models. Representative company: Databricks

From AI Hot Storage to Decentralized Cold Memory: Efficiency Maximization vs. Trustworthiness Maximization

AI storage is an extreme efficiency-driven system, its value function focuses on maximizing compute output. HBM bandwidth determines whether the GPU can be fed full, SSD throughput determines dataset and Checkpoint read/write efficiency, low latency concerns the real-time experience of RAG and inference. These indicators ultimately converge into GPU utilization and unit Token cost, directly determining the commercial profit and loss of AI applications. The ultimate goal of AI storage is not preservation, but acceleration, serving productivity.

Whereas the value function of decentralized storage is completely different. It asks whether data still exists 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 a public data base that is open access and permanently preserved. Its ultimate goal is to defend the absolute truth and sovereignty independence of data, serving fairness, censorship resistance needs, and civilization memory.

AI storage is "hot storage" that provides fuel for future productivity, decentralized storage is "cold memory" that preserves 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 the model runs, the latter determines whether memory will be deleted. Currently, market mechanisms reward the efficiency of productivity, AI storage is at the forefront, while decentralized storage seems to be experiencing the silence of valuation collapse and narrative bloodletting.

Vision and Reality of Decentralized Storage

There are many decentralized storage projects, but according to industry mindshare and ecological precipitation, the core representatives are always Filecoin and Arweave. Although both belong to "decentralized storage," their underlying architectural philosophies are almost two completely different paths—the former approaches the elasticity of AWS with market contracts, the latter approaches the eternity of libraries with one-time social contracts.

• Filecoin: Built the most complete verifiable economic system through PoRep and PoSt. It should not continue to fight hard with AWS on consumer-grade network disks, but should turn to AI data provenance (Provenance), public dataset hosting, and compliance archiving, providing a verifiable chain for model auditing and copyright proof. The necessary path is to encapsulate as S3 compatible API and support fiat payment, upgrading from "cheap storage market" to "verifiable compute infrastructure."
• Arweave: With the narrative of "pay once, store forever," through Blockweave and SPoRA mechanisms, it forces incentives for miners to save and quickly access as much historical data as possible, especially scarce historical data. Its best position is the human public memory base—saving human rights records, war crimes evidence, cultural classics, archiving legal and financial history, providing permanently accessible long-term memory for AI Agents. Arweave's value lies not in speed, but in the capacity to carry civilization memory across cycles.

The dilemma of decentralized storage projects such as Filecoin and Arweave does not lie in wrong value propositions, but in the long-term mismatch of productization, retrieval experience, real demand, and Token incentives. This reveals the huge gap from geek concepts to mainstream commercial applications:

• Supply-Demand Incentive Mismatch: Early networks represented by Filecoin expanded capacity quickly through Tokens, but did not build a strong enough paid demand side, resulting in huge capacity but insufficient utilization and paid conversion. Rewarding "I can store," rather than "Need me to store."
• Lack of Enterprise-Level Service Capability: AWS's barrier is not hard drives, but the "data operating system" composed of API, SLA, permission management, compliance auditing, and technical support. Enterprises buy "peace of mind," not experimental infrastructure that requires handling keys and node selection themselves.
• Retrieval Experience Shortcoming: "Storing in" does not equal "retrieving out stably and with low latency." Scattered nodes, complex topology, lack of unified SLA, make it difficult to undertake AI hot data workflows, more suitable for trustworthy cold archiving and data provenance.
• Insufficient Privacy Compliance: Enterprise private data cannot be simply written to public permanent networks; there is a natural conflict between the right to be deleted and permanent immutability. Decentralized storage is more suitable for public data and long-term archives, rather than indiscriminately undertaking core private data.
• Token Economy Amplifies Cycles: Bull market financialization masks insufficient demand, bear market miner ROI decline exposes commercial shortcomings. Tokens can cold-start supply, but cannot automatically create demand and sustainable revenue.

And other decentralized storage projects mostly focus on specific ecosystems or niche tracks: Storj/Sia cross-cycle industry mindshare and Web3 narrative influence are weaker than Filecoin / Arweave; BNB Greenfield/Walrus bind to specific public chain ecosystems of BNB or SUI; Celestia/EigenDA etc. belong to the Data Availability (DA) layer, serving Rollup transaction confirmation rather than long-term archiving; 0G etc. AI / DA hybrid narrative projects attempt to integrate storage, data availability, compute, and AI agent settlement into a set of AI-native modular infrastructure, but their real demand, developer adoption, and commercial closed loop remain to be verified.

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

In the period of technology dividend explosion, capital crazily chases efficiency, assets such as GPU, HBM are given extremely high premiums, decentralized storage advocating "trustworthiness and fairness" is naturally marginalized. However, the pendulum of history will not stay at the efficiency end forever. Unreasonable bans and content deletion by super platforms, AI copyright litigation outbreaks forcing data source proof, data sovereignty struggles triggered by geopolitical conflicts, disappearance of public archives caused by data monopolies, and audit pressure from regulators on model training data compliance etc. events may all brew a repricing of "trustworthy storage," and the future opportunities of decentralized storage still have a chance to reflect unique value in the following directions:

• AI Data Provenance: Combine cryptographic proofs to build "data lineage proof," coping with regulatory and audit pressure.
• Public Datasets and Civilization Archives: Anchor censored archives and cultural heritage, building irreplaceable undeletable memory.
• Trustworthy Archiving and Compliance Evidence: Achieve trustworthy self-proof through Hash evidence, providing high-level digital notarization.
• ZK/TEE/DID Technology Integration: Resolve privacy tension, upgrading from a single "storage protocol" to "trustworthy data infrastructure."
• Invisible Product Route: Provide S3 compatible API and fiat billing, allowing users to directly purchase "trustworthy archiving" services.

AI storage and decentralized storage, one pursues extreme efficiency, providing fuel for us to run towards the future; the other defends silent memory, guarding our right to look back at the past. The current market rewards efficiency without reservation, decentralized storage therefore appears silent or even collapsed; but when the AI era further amplifies data monopolies, copyright disputes, and the fragility of historical memory, decentralized storage may welcome value revaluation in the posture of "trustworthy cold layer." Those memories that cannot be easily erased by platforms, companies, or any single power, may perhaps change from idealistic romance, from marginal belief to necessary infrastructure.

Disclaimer: This article was assisted by AI tools such as Claude Opus 4.8, ChatGPT-5.5, Qwen 3.7 during the creation process. The author has tried best to proofread and ensure information is true and accurate, but omissions are still inevitable, please understand. It needs to be specially noted that the content of this article is only for information integration and academic / research exchange, does not constitute any investment advice, nor should it be regarded as any token buying or selling recommendation.