IOSG author Jacob Zhao said in an article republished by WuBlockchain that the AI era is repricing storage, but the revaluation is centered on hot-data efficiency, while decentralized storage is still built around trusted cold data. The article argues that both sectors carry the word “storage,” yet they serve different value functions.
The piece notes that CXMT, described there as China’s first listed storage stock, officially debuted on ChiNext last month and surged 500%, drawing broad market attention. Even with the storage sector still dealing with a pullback, AI storage remains a key part of the current technology narrative in capital markets. Web3 decentralized storage, on the other hand, has stayed quiet for a long stretch.
Why storage has moved back to the center of the AI stack
In the traditional IT era, the article says, storage was mainly a capacity business. Enterprise CIOs looked at cost per unit, drive reliability, disaster recovery, archiving policies, and hardware refresh cycles that often ran 3–5 years. Storage was usually treated as something attached to server procurement.
IOSG argues that the current rally is not a standard cyclical recovery. It is a repricing of data movement under the AI model era. Storage logic has shifted from “capacity first” to “efficiency first,” with attention now on GPU feed rates, checkpoint writes, and very low latency for retrieval-augmented generation, or RAG. In that framing, storage is no longer just the final resting place for data. It becomes a high-speed path that brings data into computation.
The article describes the evolution of AI infrastructure bottlenecks as a battle against the bucket effect. Real compute utilization is presented not as a linear sum of assets but as a multiplier: GPU × HBM × DRAM × SSD × network × file system. A weakness in any one layer can drag down the whole system. That, IOSG says, is the basic reason storage has shifted from a cost center to an efficiency engine in the AI era.
The AI storage stack, from memory close to compute to the data lake foundation
IOSG breaks the AI storage architecture into four core layers and says market value and investor attention are concentrated around HBM, enterprise SSDs, SSD controllers, NVMe and CXL protocols, and high-performance storage systems.
Memory close to compute
This layer is led by HBM, with DRAM and CXL-based memory pooling as supporting parts. It sits close to GPU or CPU packages and buses, with the goal of breaking the memory wall. The article calls it the first checkpoint in whether compute can be fully released.
High-speed persistent storage
The core logic here is defined as: enterprise SSD = NAND + SSD controller + NVMe/PCIe data path. According to IOSG, this is the layer that handles frequent checkpoint writes, large-scale training set loads, and hot-data caching for RAG. It is also described as the clearest source of persistent storage growth inside AI data centers.
Low-cost, large-capacity storage
This level includes HDDs, cold storage, and data lake archiving systems. It remains important because multimodal raw data, historical logs, and compliance backups are all expanding quickly. IOSG says this segment still holds an irreplaceable total cost of ownership advantage.
Storage systems and data software
This group includes high-performance parallel file systems, distributed object storage, vector databases, and the RAG data governance layer. The point, the article says, is that AI does not really consume bare hardware. It consumes data availability after software has organized, indexed, and permissioned the data.
Outside those hot-data layers, the article places decentralized storage in a separate position as a trusted cold layer. Rather than competing in millisecond-level AI data races, it is aimed at public dataset attestation, AI training data provenance, and long-term archival memory.
HBM, DRAM, CXL, and enterprise SSDs each solve a different problem
HBM as the bandwidth organ closest to compute
IOSG says high-bandwidth memory, or HBM, is not storage in the traditional sense. It is a high-bandwidth memory tier located close to the GPU. Its job is not to preserve data, but to keep feeding data into compute at very high bandwidth. The article calls it the most deterministic part of the AI storage chain and one of the clearest supply bottlenecks.
Its architecture is summarized as “3D DRAM stacking + 2.5D advanced packaging.” With TSV vertical stacking and CoWoS heterogeneous integration, the distance between storage and compute is compressed to unlock a step change in bandwidth. IOSG says the moat is not only in DRAM design. It also sits in process technology, TSV, ultra-thin stacking, packaging, heat dissipation, testing, and customer certification. A yield problem in any one step can scrap an entire HBM stack.
The article says only SK hynix, Samsung, and Micron can currently produce HBM at scale on a stable basis, creating a three-part moat built on DRAM manufacturing, packaging capability, and certification from customers such as NVIDIA and AMD.
DRAM and CXL as the system memory base and memory pooling engine
HBM is used to solve extreme bandwidth near the GPU, DRAM serves as the main system memory foundation in servers, and CXL is presented as an attempt to break physical limits and reorganize memory resources across data centers.
On DRAM, the article says it mainly handles CPU-side caching, data preprocessing, temporary storage for intermediate states, and system operations. The global DRAM market is described as highly concentrated among SK hynix, Samsung, and Micron, while CXMT is identified as a key domestic substitution variable in China.
CXL, or Compute Express Link, is described as a next-generation cache-coherent interconnect protocol for data centers. Its goal is to move beyond the constraints of DIMM slots, local memory ceilings, and isolated server memory pools, pushing memory architecture toward expansion, pooling, and sharing. IOSG says CXL is still in an early stage, moving from platform support toward broader deployment, with companies such as Astera Labs and Montage Technology cited as core names.
Enterprise SSDs as the data throughput hub
The article says enterprise SSDs are the most important high-throughput persistent storage increment in AI data centers. Their role is to keep data flowing to GPUs through high throughput, low latency, and stable QoS across the full lifecycle of training data loads, checkpoint writes, RAG lookups, inference caching, and log returns.
IOSG breaks the enterprise SSD stack into three layers.
- NAND: the raw media layer that determines storage density and unit cost. The companies listed are Samsung, SK hynix (Solidigm), Micron, Kioxia, Western Digital, and YMTC.
- SSD controllers: the layer that determines performance release, error correction, QoS stability, and wear leveling. The companies listed are Phison, Silicon Motion, Marvell, and Maxio.
- NVMe/PCIe: the data path layer that determines transport efficiency from storage to compute. The article says technologies such as GPUDirect Storage can reduce CPU memory bounce buffers and CPU involvement, easing I/O bottlenecks. Companies named here include Broadcom, Marvell, and Astera Labs.
HDDs, cold storage, and archives still matter
IOSG says AI will not eliminate HDDs. Demand for low-cost cold storage is rising alongside multimodal video and image data, compliance logs, and ever larger historical datasets. In its framework, SSDs and HDDs work together in a tiered model tied to business value: SSDs handle hot data and high throughput, while HDDs handle low cost and long retention. Seagate, Western Digital, and Toshiba are listed as representative companies.
The storage software stack is what turns hardware into usable data
The article also treats storage software as the scheduling center of data availability. AI, it says, consumes not disks themselves but data services assembled through software. That software stack is divided into four layers.
- High-performance storage systems: focused on concurrency, throughput, and low latency, often through parallel file systems that address GPU cluster data starvation. The companies named are VAST Data, WEKA, and Pure Storage.
- Object storage: built around objects, keys, and metadata, carrying large volumes of unstructured data. It does not chase the lowest latency, instead leaning on low cost and cloud-native properties. AWS S3 is the example cited.
- Vector databases: used to store, index, and retrieve vectors produced by embedding models, allowing AI systems to find relevant content from large knowledge sets. The article names Pinecone and Milvus.
- RAG data layers: these go beyond retrieval alone and cover chunking, cleaning, access control, and source tracing, with the goal of making enterprise data safe, accurate, and traceable when called by large models. Databricks is cited here.
Efficiency maximization versus trust maximization
In the article’s third section, IOSG draws a direct contrast between AI storage and decentralized storage at the level of underlying purpose. AI storage is described as a system driven by efficiency, with value tied to maximizing compute output. HBM bandwidth affects whether GPUs stay fed. SSD throughput affects dataset loading and checkpoint reads and writes. Low latency shapes the real-time experience in RAG and inference. Those metrics ultimately show up in GPU utilization and cost per token.
Decentralized storage is framed differently. The article asks whether data will still exist 10 years later, whether it can be altered, and whether it can survive single-point censorship. Through cryptographic proofs and distributed networks, it is meant to build an openly accessible, permanently preserved public data base layer. In IOSG’s telling, it serves fairness, censorship resistance, and civilizational memory.
The article sums up the split this way: AI storage is hot storage built to fuel future productivity, while decentralized storage is cold memory designed to preserve records that cannot be easily erased. One is built for speed and efficiency. The other is built for trust and duration.
Filecoin and Arweave take very different paths
Among decentralized storage projects, IOSG highlights Filecoin and Arweave as the two main representatives, while stressing that their architectural philosophies diverge sharply.
Filecoin
The article says Filecoin built one of the more complete verifiable economic systems through PoRep and PoSt. In IOSG’s view, it should stop trying to fight AWS head-on in consumer cloud-drive style services and instead move toward AI data provenance, public dataset hosting, and compliant archiving. The article says that would allow it to provide a verifiable chain for model audits and copyright proof. The route it points to is S3-compatible APIs plus fiat payment support, shifting Filecoin from a cheap storage market toward verifiable compute infrastructure.
Arweave
For Arweave, the article centers on the “pay once, store forever” narrative. It says the network uses Blockweave and SPoRA to push miners to store and quickly serve as much historical data as possible, especially scarce records. IOSG places Arweave’s best use case in the role of public memory infrastructure for humanity: preserving human rights records, war crime evidence, cultural works, and legal and financial history, while also serving as long-term memory for AI agents. The article says Arweave’s value is not speed but its ability to carry civilizational memory across cycles.
The real constraints facing decentralized storage
IOSG argues that the problem for Filecoin, Arweave, and similar projects is not that their core value proposition is wrong. The issue lies in a long-running mismatch among productization, retrieval experience, real demand, and token incentives. The article lays out several concrete constraints.
Supply-demand incentive mismatch
Using Filecoin as an example, the article says early networks expanded supply quickly through token incentives without building a strong enough paying demand side. The result was huge capacity with weak utilization and poor paid conversion. In its words, the system rewarded “I can store” rather than “someone needs me to store.”
Missing enterprise-grade service layers
IOSG says AWS’s moat is not just hardware. It is the full data operating system made up of APIs, SLAs, permission management, compliance audits, and technical support. Enterprises are paying for reliability and assurance, not for experimental infrastructure that forces them to handle keys and node selection on their own.
Retrieval experience remains a weak point
The article says writing data into a network is not the same as getting it back in a stable, low-latency way. Because nodes are distributed, topology is complex, and unified SLAs are missing, decentralized storage is not well suited to hot-data AI workflows. It fits trusted cold archiving and data provenance better.
Privacy and compliance tensions
IOSG notes that private enterprise data cannot simply be written into a public permanent network. The right to deletion can clash with permanent immutability. That makes decentralized storage better suited for public data and long-term archives than for indiscriminately handling core private datasets.
Token economics can magnify the cycle
According to the article, financialization in bull markets can hide weak real demand, while falling miner ROI in bear markets exposes commercialization problems. Tokens can help cold-start supply, but they do not automatically create demand or durable revenue.
Other projects and adjacent sectors
Beyond Filecoin and Arweave, the article also sorts several other projects and sectors by role.
- Storj and Sia: described as weaker than Filecoin and Arweave in long-cycle industry mindshare and Web3 narrative impact.
- BNB Greenfield and Walrus: presented as tied to the specific chain ecosystems of BNB and SUI.
- Celestia and EigenDA: categorized as data availability layers for Rollup transaction confirmation rather than long-term archiving.
- 0G: described as trying to combine storage, data availability, compute, and AI agent settlement into an AI-native modular infrastructure stack, though the article says real demand, developer adoption, and a commercial loop still need verification.
Where IOSG sees future openings
In the final section, IOSG says capital tends to chase efficiency during periods of strong technical upside, giving assets such as GPUs and HBM rich premiums and leaving decentralized storage on the margins. Still, the article argues that the pendulum will not stay at one end forever. It says storage built around trust could be repriced if pressure rises from platform bans and content removal, AI copyright litigation, geopolitical fights over data sovereignty, the disappearance of public archives, or compliance audits around training data.
The article lists several areas where decentralized storage may still show distinct value.
- AI data provenance: using cryptographic proofs to build a verifiable lineage trail for regulatory and audit needs.
- Public datasets and civilizational archives: focusing on censored records and cultural heritage to build memory that cannot easily be deleted.
- Trusted archives and compliance attestation: using hash-based proofs to support trusted self-verification and high-level digital notarization.
- Integration with ZK, TEE, and DID: reducing privacy tensions and expanding from a storage protocol into a trusted data infrastructure layer.
- Invisible product paths: offering S3-compatible APIs and fiat billing so users buy a trusted archive service directly.
IOSG closes by drawing a final split. AI storage is chasing the limits of efficiency and fueling future productivity. Decentralized storage protects memory that should not be easily erased. The market is rewarding efficiency today, the article says, but if AI intensifies concerns around data monopolies, copyright disputes, and the fragility of historical memory, decentralized storage may return as a trusted cold layer with a renewed valuation case.