AI storage opportunity brings new risks for enterprise vendors

The AI storage opportunity threat is now visible across enterprise infrastructure: AI workloads are increasing demand for fast data access, larger protected data estates and tighter governance, while AI agents create new ways for data to be changed, misused or attacked. For storage vendors, the shift has moved the market away from a steadier enterprise array cycle and toward systems built around GPUs, parallel access and AI-ready data pipelines.

ChatGPT’s release 44 months ago marked the point at which storage suppliers had to respond to a different demand curve. AI training first drove much of the pressure, but enterprise inference is taking a larger role as companies build so-called AI factories to tune, deploy and run internal and third-party agents.

How is AI changing storage?

AI systems need to pull large volumes of structured and unstructured data into GPUs without leaving expensive processors waiting on input/output. That requirement has pushed suppliers toward SSDs, NVMe and PCIe, replacing older disk-era SAS and SATA designs in performance-sensitive systems.

Nvidia has supported storage technologies aimed at keeping GPUs fed. GPUDirect reduces the need to move data through storage controllers and host memory before it reaches GPU systems, while RDMA-based access has been applied to file and object storage. Cloudian, MinIO and Scality have worked in S3-over-RDMA-style object approaches, according to the storage sector’s current vendor positioning.

KV caching is also being used to extend the effective memory available to AI workloads by moving some data pressure from GPU high-bandwidth memory to SSDs, reducing waits and avoiding repeated token computation. Nvidia has partnered with major enterprise storage vendors on these approaches.

The architectural split is becoming clearer. Disaggregated storage array technology, associated with VAST Data, has been adopted by Dell, HPE, NetApp and Everpure. High-end monolithic array suppliers such as Hitachi Vantara, IBM and Lenovo-owned Infinidat have not adopted the same DASE, GPUDirect or KV caching pattern, leaving them more tied to legacy enterprise roles.

Which vendors are gaining from the shift?

AI has helped newer storage and data infrastructure suppliers compete with incumbents. DDN, Pure Storage, VAST Data and WEKA have grown with demand for parallel data access, flash-based systems and high-performance storage tied to AI and analytics. Databricks and Snowflake have become central AI data lake platforms, while GPU-as-a-Service providers such as CoreWeave and Lambda have created a cloud market built around scarce accelerator capacity.

Vector data has created another layer of demand. Dedicated vector database companies including Pinecone, Qdrant, Weaviate and Zilliz serve AI embedding search, while broader database platforms such as SingleStore have added vector support.

Data management vendors including Arcitecta, Datadobi, Hammerspace and Komprise are positioning around classification, metadata access and selective data movement. Apache Iceberg is being used to connect data lakes with external storage without forcing every dataset into one physical repository.

Why AI agents create a storage security problem

AI agents can access, modify and move data inside IT systems at machine speed. That makes agent identity and access management a storage issue as well as a security issue, because agent actions need to be governed, logged and reversible when mistakes or malicious activity occur.

Backup and cyber-resilience vendors are extending their roles accordingly. Cohesity, Commvault, Druva, Rubrik, Veeam and others are using backup data as a source for AI search, recovery analysis and attack detection. Cohesity’s Gaia is one example of a generative AI search assistant built around protected data.

These vendors are also promoting the use of AI to identify the start of an attack, locate affected data, find the last clean copy and support recovery. Cohesity, Druva, HYCU, Rubrik and Veeam are active in agent governance, using AI systems to monitor other AI systems.

Storage vendors have used machine learning in operations for years, including telemetry monitoring popularized by HPE-acquired Nimble. The newer step is an AI interface for administrators, allowing staff to query protection gaps, policies and fleet behavior in natural language. That may reduce manual work, but it also raises the same control issue facing the broader AI stack: storage teams will need trusted agents to supervise other agents.

This story draws on original reporting from The Register.