At the IBM Think conference in May 2026, the atmosphere was electric with the promise of generative AI, but beneath the surface-level excitement, a more sober conversation was happening regarding the physical and logical foundations of this new era. Sam Werner, the general manager who oversees the development of the entire IBM Storage portfolio, and Christopher Vollmar, a Master Inventor and global architect focused on operational resiliency, are the minds tasked with solving the friction between rapid AI adoption and the rigid requirements of enterprise storage. They have seen firsthand how organizations often treat data as a static resource to be moved at will, failing to realize that AI workloads demand a fundamentally different approach to security, governance, and infrastructure. This discussion explores the often-overlooked vulnerabilities of vector databases, the necessity of content-aware storage, and why the traditional method of copying source data is becoming a liability in a world governed by strict data privacy and zero-trust principles.
Many organizations currently rely on copying source data into vector databases for retrieval-augmented generation, yet these copies often remain disconnected from the source. What are the specific governance and security risks that IT leaders are missing when they take this shortcut?
When IT teams decide to simply move files around to feed a vector database, they are often telling themselves a “good story” about AI readiness that doesn’t actually hold up under the pressure of a real production environment. The moment you copy a file from its original location to vectorize it for RAG, you have created a ghost of that data that is totally untethered from its origin. This creates a terrifying gap in governance: if a source file is updated because it contains sensitive PII or if a data retention policy dictates it must be deleted, that change doesn’t ripple through to the vector database. You end up with a searchable, discoverable index that contains information you are no longer legally allowed to possess or share. This oversimplification is a major blind spot because it ignores the foundation of enterprise storage, which is built to manage access rights and controls meticulously; by bypassing that foundation, you leave your most valuable corporate data vulnerable on unprotected servers. It is a risky gamble that treats data like a static image rather than a living, breathing asset that requires constant oversight.
You’ve mentioned that there are four critical dimensions—distributed, diverse, dynamic, and dark data—that define a modern AI data plane. How should organizations rethink their storage architecture to address these specific dimensions?
Building a functional AI data plane requires moving beyond the idea of a central repository and instead focusing on an architecture that can handle the complexity of the four emerging standards. First, you have the distributed nature of data, which must be abstracted so that AI systems can pull from multiple locations without the need for constant, manual re-platforming of every single byte. Then there is the diversity of the data itself, which might exist as a file, an object, or a vector and needs to be accessed through various protocols and APIs simultaneously. We also have to consider dynamic data, which requires a vertical acceleration from the source directly into AI system memories to keep pace with the speed of modern inference models. Finally, the most dangerous category is “dark data”—the massive piles of unclassified information that sit in the shadows of an organization like a library where the lights have been turned off and the catalog has been lost. This dark data needs to be classified, governed, and cataloged for awareness by AI systems, turning a potential liability into a structured asset that storage can help manage through content-aware technologies.
As AI workloads become more complex, the attack surface for enterprise data seems to expand exponentially. How can storage infrastructure provide a layer of protection that goes beyond traditional software-level security?
The current trend of copying data to various servers for AI processing is an invitation for disaster because it strips away the protective discipline typically found in high-end storage arrays. To combat this, we have to bring that data back under the umbrella of enterprise-grade security, which includes mandatory encryption of data both while it is in flight and while it is resting on the physical disks. This isn’t just about locking a digital door; it’s about a comprehensive approach to data governance where we are constantly tracking every access right and control to ensure that only those with explicit permission can see the information. When you rely on a solid storage foundation, you are using the array’s internal intelligence to maintain the sanctity of the data, preventing the messy sprawl that happens when researchers or developers spin up isolated vector databases without the oversight of the IT security team. By keeping the data under the foundation of a managed enterprise array, you ensure that the same security standards applied to your financial records are also applied to the data fueling your LLMs, creating a hardened shell around the most sensitive parts of the business.
In the context of operational resiliency, what specific actions should a company take to move toward a zero-trust model for their storage and AI pipelines?
Moving toward zero trust is a rigorous process that demands diligence across every single layer of the stack, from the compute level down through to the deepest backup and archive layers. One of the most critical actions an organization can take is the implementation of two-person integrity, which acts like the dual-key system on a submarine, ensuring that no single individual has the power to destroy or alter critical data sets without oversight. We also push for the adoption of multi-factor authentication for administrative access and a complete re-evaluation of the roles assigned to people who manage the infrastructure. Beyond just human controls, the storage itself must be used to create immutable copies of the data—backups that are functionally frozen in time and cannot be changed or deleted by ransomware or accidental errors. These immutable copies allow for a rapid recovery and proactive validation of the data throughout its entire lifecycle, ensuring that the business can keep operating even when the primary systems are under threat or being audited for compliance.
Storage is often seen as a passive container for information, but you’ve talked about “content-aware storage” as a way to increase the value of AI. How does this technology actually bridge the gap between raw data and actionable intelligence?
The real value of AI doesn’t come from a generic model; it comes from combining that model’s reasoning intelligence with the unique, timely data found within a corporation’s own walls. Content-aware storage acts as a bridge by helping to manage the semantics of the data and automatically cataloging metadata, which essentially tells the AI system exactly what it is looking at and why it matters in a real-world context. Instead of a manual, labor-intensive process of preparing data for AI, storage systems can now help automate the vectorization of corporate information for similarity searches. This means the storage is no longer just holding bits and bytes in a silent box; it is actively helping the AI to understand the context of the information, which leads to much more accurate and relevant inferencing. By managing these semantics at the storage level, we are taking a massive burden off the data scientists and ensuring that the corporate intelligence being fed into the AI is as fresh and organized as possible, turning the storage array into an active participant in the AI reasoning process.
What is your forecast for AI storage?
In the coming years, we will see a shift away from storage as a silo and toward a unified AI data plane where the lines between data management and AI development completely disappear. Organizations will stop seeing storage as an afterthought and start viewing it as the primary engine for AI performance and security. We will likely see a massive move toward zero-trust data foundations, where encryption and immutable copies are not just features but the standard operating procedure for any company that wants to avoid the legal and financial ruins of a data breach. The era of just copying files to see if the AI works is coming to a close, and it will be replaced by a sophisticated architecture that respects data gravity, ensures total governance, and uses content-aware technology to turn every byte of a company’s history into a competitive advantage. Success will be defined by those who can unify their distributed, diverse, and dynamic data into a single, secure, and highly accelerated flow.
