Storage & Data Management Watch · October 6, 2026
NetApp, VDURA, Oracle, Jedify, and Cloudflare are all answering the same question: when AI agents consume data at machine speed, who prepares, governs, and protects it?
Core Research Question
Is AI turning storage and data management from a passive repository into an active intelligence layer that prepares, governs, and protects data for models and agents?
1. The Data Infrastructure Shift
Storage used to be the quiet layer under the stack: capacity, throughput, and a backup schedule. This week's announcements, many from NetApp INSIGHT 2026, suggest a different role. Storage and data management are being asked to prepare data for models, police what AI agents can do with it, and keep expensive GPUs supplied.
The seven developments below cover performance, metadata, governance, agents, memory economics, cloud data platforms and protection. The common thread is that the data layer is becoming active rather than passive.
The CODEW angle: The next AI bottleneck may not be compute alone. It is the ability to organize, govern, move, protect, and continuously serve the data AI systems consume.
2. NetApp Novus: AI Data Infrastructure Moves Toward Zettabyte Scale
NetApp is positioning its Novus architecture around the metadata bottlenecks that large GPU clusters create. According to SiliconANGLE's INSIGHT coverage, the company says Novus exceeds 100 TB/s of aggregate throughput, and its Data Director manages metadata separately from the stored data so each can scale independently. It also gives applications a unified view of files across multiple ONTAP clusters.
Speed is only half the pitch. NetApp's AI Data Engine discovers, classifies, and vectorizes enterprise data, aiming to cut data-preparation work that NetApp's chief product officer says typically takes six to nine months of engineering. NetApp has also announced its intent to acquire PEAK: AIO, which it expects to bring parallel file system technology built on the pNFS protocol, so customers can add AI infrastructure without replacing existing systems.
The CODEW angle: Storage is no longer just where enterprise data lives. It increasingly determines whether that data can be used by AI at all, and NetApp's performance figures are company claims.
3. VDURA V12 Brings the Hyperscaler Storage Model to AI Factories
VDURA announced general availability of its Data Platform V12, which it describes as turning the platform into a multi-tenant, API-driven storage service for GPU clouds and AI factories. V12 combines file and S3 object access in one platform, adds RDMA data paths that let GPUs exchange data with storage without routing through the CPU, and is demonstrated on Supermicro hardware.
The pitch is explicitly about neoclouds: storage determines how much rented GPU capacity turns into revenue, so tenants must be isolated, cold data tiered cheaply, and warm data moved back to flash quickly. VDURA's CEO says every kilowatt spent on storage is a kilowatt not spent on compute, tying storage design to the power constraint our Cloud Computing Watch tracks. Independent analysis notes that similar distributed-metadata approaches already exist at IBM, WEKA and VAST Data.
The CODEW angle: Parallel file systems, GPU infrastructure and data management are converging. In neocloud economics, storage efficiency is GPU utilization by another name.
4. AI Agents Create a New Data-Governance Problem
AI agents touch data at machine speed, and much of the risk comes from unintended behavior rather than malicious intent. NetApp's Russell Fishman told theCUBE that the company is moving access control away from simple yes-or-no permissions toward behavior-based policies, with a new engine in AI Data Engine that flags unusual patterns of data use. His example: an agent reading one customer account is normal, while an agent reading all of them is not.
The same data layer is also being tied into cyber resilience. NetApp's autonomous ransomware protection now feeds signals into recovery workflows from Commvault, with similar links to Microsoft Sentinel and Splunk. NetApp says AI Data Engine can be deployed against any existing NetApp system and announced a free six-month offer.
The key question is who controls what an agent can see, retrieve, modify,, and distribute, a theme we also follow in Cybersecurity Watch.
The CODEW angle: Governance is moving from who has access to how access is being used. Storage vendors are positioning themselves as an enforcement point for agent behavior.
5. Enterprise Data Governance Becomes an AI-Readiness Requirement
An Oracle Cloud Infrastructure blog published October 2 describes a governance framework on Oracle AI Data Platform. Note that it is a custom approach written by an Oracle consultant, not a new product release. It combines AI-assisted discovery of sensitive fields, steward approval, configurable masking, role-based access control, and continuous checks that deployed controls still match approved policy.
The design keeps humans in control: recommendations carry a confidence score and plain-language reason, only a steward's committed decisions become policy, and the governance agent explains verified findings without approving classifications, applying masking or changing permissions itself.
The CODEW angle: The practical lesson is auditability. Discovery, classification, access controls, and drift detection form one lifecycle, and AI is allowed to recommend rather than decide.
6. Data Agents Turn Data Management Into an Active Workload
Jedify launched Data Agents, which monitor business metrics, investigate anomalies, and run recurring analysis on triggers or schedules, built on its context graph that maps entities, definitions, and permissions across fragmented sources. Agents inherit the requesting user's existing data permissions, so they cannot reveal rows, columns, or metrics the user could not see, and they are exposed through MCP as tools callable from Claude, ChatGPT, Slack, or other agent platforms.
Blocks & Files notes that agent builders are table stakes in enterprise software but rare in storage and data protection, where Commvault is one of the few with one. Jedify, which raised a $24 million Series A in June, claims its pre-resolved graph uses far fewer tokens than agents that infer meaning from raw schema.
The CODEW angle: The next generation of data platforms may not only store and organize data. They may continuously interpret it and act on it.
7. Memory Has Become a Massive Part of the AI Infrastructure Economy
Blocks & Files tabulated the last four quarters of revenue for Kioxia, Micron, Samsung, SanDisk and SK hynix and arrived at $483.7 billion, nearly four times the prior level, as AI servers pulled in HBM-heavy GPUs, DRAM-loaded CPUs and all-flash storage. HBM supplier revenue accounts for the largest share, followed by DRAM and NAND. The publication's own extrapolation, which is an estimate rather than a forecast from the companies, suggests another $1.5 trillion could flow to these five suppliers by 2030 if the imbalance persists.
For data infrastructure, the consequence is cost. The same boom raises demand for HBM, DRAM, NAND, enterprise SSDs and raw storage capacity at once, and The Register reported this week that Nvidia shipped a new $4,999 DGX Spark with half the RAM and storage amid the memory crunch. Chip-manufacturing detail belongs in Semiconductor Watch; the storage question is how enterprises budget for data-heavy AI when flash and memory stay expensive.
The CODEW angle: Memory is no longer a component line item. It is a budget category that shapes how much data enterprises can afford to keep hot.
8. Cloud Data Platforms Move Closer to the Storage Layer
Cloudflare moved its Data Platform into general availability under the name Basin: Pipelines, R2 Data Catalog and R2 SQL become Basin Pipelines, Basin Catalog and Basin SQL. Built on Apache Iceberg, the platform stores data in R2 and lets customers query it with Basin SQL or external engines including DuckDB, Snowflake, Spark and Trino.
Cloudflare's pitch is no lock-in and no egress fees. The Register points out the nuance: egressing directly from R2 is free, but Basin Pipelines charges for delivering data to its sinks after an included allowance, and other metered services connected to R2 may add costs.
The CODEW angle: The larger question is whether cloud data is consolidating into integrated storage, catalog, processing, and analytics platforms. Open table formats such as Iceberg make that consolidation less risky for buyers.
9. The Data Infrastructure Layer Map
Seven layers now compete to make data usable for AI. Today's developments sit in each of them.
| Performance | Metadata scaling, parallel file systems and RDMA keep GPUs fed (NetApp Novus, VDURA V12). |
| Preparation | Discovery, classification and vectorization make data usable by models (NetApp AI Data Engine). |
| Governance | Behavior-based policy, steward approval, masking and drift checks control agent access (NetApp, Oracle). |
| Data agents | Context graphs and MCP turn analysis into continuous, permission-aware workloads (Jedify). |
| Media economics | HBM, DRAM and NAND scarcity raises the cost of every tier ($483.7B memory revenue). |
| Cloud platform | Open-format storage, catalog and SQL converge in one service (Cloudflare Basin). |
| Protection | Storage signals feed ransomware detection and recovery workflows (NetApp with Commvault). |
The CODEW angle: No single layer wins on its own. Vendors that combine performance, governance and protection are making the strongest case that storage is becoming an intelligence layer.
The CODEW Angle
AI is turning storage and data management into an active intelligence layer.
The strongest developments this week are not about capacity. They are about metadata, governance, agent access, memory economics, and the infrastructure needed to feed production AI. Each of them asks the data layer to understand, control, and respond to how data is used.
The next bottleneck in AI may not be compute alone. It is the ability to organize, govern, move, protect, and continuously serve the enormous volumes of data AI systems require.
Sources
→ SiliconANGLE: NetApp's Novus tackles metadata bottlenecks in AI data infrastructure
→ SiliconANGLE: Storage admins get automated guardrails as agents outrun human-in-the-loop review
→ HPCwire / BigDATAwire: VDURA V12 brings the hyperscaler storage playbook to AI factories
→ SDxCentral: VDURA storage platform update targets neoclouds
→ NAND Research: VDURA Data Platform V12 research note
→ Oracle: Building an AI-assisted sensitive data governance framework on Oracle AI Data Platform
→ Blocks & Files: Jedify customers can build agents to use its context graphs
→ Blocks & Files: Almost half a trillion dollars sucked up by memory industry
→ The Register: Cloudflare launches Data Platform with bland Basin branding
→ The Register: Nvidia debuts $4,999 DGX Spark with half the RAM and storage
The CODEW Stat
8 developments · $483.7B memory-supplier revenue in four quarters · 100+ TB/s claimed (NetApp Novus) · 7 layers of the AI data stack
Performance claims are vendor figures; the memory total is a Blocks & Files tabulation of five suppliers' reported revenue.
Editorial Note
Storage & Data Management Watch is the fast-moving intelligence layer tracking what is changing now in AI storage, enterprise data platforms, governance, memory economics, cloud data services and data protection, distinct from Semiconductor Watch (chips and fabs), Hardware Watch (devices and servers) and AI Watch (models and agents).
This edition examines whether the data layer is becoming an active intelligence layer, drawing on SiliconANGLE, Blocks & Files, The Register, Oracle and company announcements as of October 6, 2026.
Educational content only. Not investment advice. Performance and efficiency figures are vendor claims and have not been independently verified by The CODEW. SiliconANGLE and theCUBE are paid media partners for NetApp INSIGHT; their event coverage is cited as reported. The memory revenue total is a Blocks & Files tabulation, and projections are that publication's estimates.
Reviewed by Erwin Castro
on
Tuesday, October 06, 2026
Rating:
No comments: