Data has long been described as the foundation of AI. That part is not new. What is changing is the role data plays.
In a recent presentation at the AI-Ready Data Summit, Himabindu Tummala, Distinguished Engineer at Dell Technologies, showed how AI is evolving right now and what that evolution demands from the data layer.
As AI systems mature, data is no longer just static input. It becomes context, memory, and state. Models reason over historical information, retrieve embeddings in real time, and maintain long-lived context across tasks. That is why the data platform, especially its core storage layer, is becoming increasingly central to how AI systems are built and scaled.
AI Evolution Is Raising the Bar for Data
Over the last few years, each phase of AI has introduced new pressure on data systems:
The generative AI wave pushed large-scale training, open models, checkpointing, and the need to move massive volumes of data efficiently.
The shift toward agentic AI raised the stakes further because models are now expected to reason, plan, and act. That makes fast inference and real-time data access much more important.
The next stage, often described as physical AI, raises the bar again with reinforcement learning, large agent ecosystems, and much stricter expectations around security, resilience, and operational reliability.
What matters most is that this progression is cumulative, not disruptive. Each phase builds on the previous one.
Analytics led to generative AI. Generative AI is leading into agentic systems. Agentic systems are laying the groundwork for physical AI. That means the architectural choices organizations make today will strongly influence how practical and efficient the next steps become.
As teams move along this curve, the same technical pressure points recur. Data volumes keep growing. Latency and throughput become more sensitive. Governance, security, and reliability move from being helpful to being mandatory. At that point, simple storage is no longer enough. A real AI data platform architecture becomes necessary.
The Shift From GenAI to Agentic AI
The move from generative AI to agentic AI is one of the most important transition points.
Agentic systems introduce persistent context, longer-lived memory, and action over time. That fundamentally changes the role of the data layer. The key questions become: where is context stored, how is it retrieved, how efficiently can agents access historical and operational data, and how is all of that governed over time?
These are central architecture questions, not side details.
As AI becomes more autonomous, friction in the data layer becomes dramatically more expensive. A small inefficiency in a traditional analytics workflow might be tolerable. The same inefficiency in an agentic system can cascade into slower reasoning, higher cost, lower reliability, and worse outcomes. Data readiness today directly shapes AI capability tomorrow.
The goal is not to guess exactly what comes next. The goal is to build a data architecture that is open, scalable, secure, and resilient enough to evolve with whatever comes next.
The Human Loop Behind AI Success
Technology is only part of the story. The other part is the teams that make these systems work.
Across modern data environments, three roles repeatedly shape the success of AI initiatives:
- Data engineers who prepare and operationalize data
- Data scientists who use it to build and refine models
- Data analysts who turn outputs into insights and decisions
What matters is not just the existence of these roles, but the continuous loop between them.
All of this runs on a shared foundation of data pipelines, transformations, and increasingly multimodal data structures. Structured and unstructured data now sit side by side. As generative AI and agentic systems mature, unstructured data ceases to be secondary and becomes first-class.
That has changed the expectations on the platform underneath. Data is no longer treated as a one-off pipeline or a temporary project asset. It’s expected to become a reusable, trusted, governed, and performant data product.
From a technology perspective, that means engineers need platforms that can ingest and transform data reliably, scientists need fast and consistent access, and analysts need governed information they can trust. All of that has to happen without creating brittle, custom-built architectures that constantly need rework.
This is where AI-ready data platforms become essential. They must support shared access patterns, mixed workloads, structured and unstructured data, and different performance profiles, all in the same environment.
Modern AI success comes from reducing friction across the whole data team, not from optimizing for one persona at a time, so insights, models, and agents can evolve together.
The AI Data Lifecycle
If we zoom in on how teams actually work, a useful way to think about the problem is as a continuous lifecycle.
Data discovery
Before teams can build anything useful, they need to know what data exists and where it lives. That challenge grows harder as datasets expand and become more distributed.
Data curation
Once the data is identified, it has to be curated. With structured data, that usually means combining disparate sources and making them usable for downstream systems. Increasingly, this goes beyond traditional SQL workflows. With unstructured data, the challenge is larger. Documents need to be parsed, meaning extracted, embeddings created, and raw content transformed into something that models and agents can actually use.
At that point, metadata, lineage, and transformation pipelines stop being optional support functions and become first-class architectural concerns.
Data retrieval
Teams need to identify what data is actually relevant for a given task. That means search becomes a platform-level responsibility. Different workloads require different search methods and different service-level expectations. The platform must handle indexing, access enforcement, and multiple search patterns consistently and efficiently.
Data preparation
Once the right data has been found, teams need to produce a usable subset for training or fine-tuning, rather than exposing every available source. That matters for security, efficiency, and cost. This stage is extremely sensitive to latency, throughput, and data locality, because friction here directly slows training cycles and raises cost.
Monitoring
Finally, the lifecycle doesn’t end at access. It has to be continuously monitored and optimized. As models, users, and data all grow, teams need to understand usage patterns, manage growth, and optimize performance and storage cost while preserving governance and resilience.
The broader point: AI calls for a continuous data lifecycle, not a collection of isolated tools.
As systems become more autonomous, data platforms have to evolve from passive repositories into active systems that support reasoning, learning, and action at scale.
What AI Agents Change
Because so much of today’s conversation centers on agentic AI, it is worth pausing on what agents actually are.
At a high level, agents are software systems built from a few core pieces. They include:
- A model, which gives the system the ability to observe, reason, and plan
- A system prompt that defines the agent’s purpose and provides higher-level instructions
- Tools that enable the agent to interact with the outside world, such as retrieving data, calling services, or sending requests to other systems
- Memory, both short- and long-term, which allows the agent to learn from past actions and improve future performance
These elements are tied together by orchestration logic, often implemented through software frameworks that help coordinate planning, tool use, and memory. On top of that sits an interface or API through which the agent receives requests from users, systems, or other agents.
Together, these components create something more than a model that simply responds. They create a system that can reason, act, remember, and pursue goals over time.
Scaling Agents Means Scaling Shared Infrastructure
Although people often talk about agents as single entities, in practice, they’re usually distributed systems made of shared services.
Two agents may serve completely different purposes. One might handle customer support escalation while another processes expense reports. They may have different prompts, different tools, and different memory contexts. But under the hood, they may share the same inference service and the same core infrastructure.
That matters because it changes the economics of scale.
Instead of deploying a separate language model for each agent, organizations can operate a single powerful inference service that handles reasoning across many agents. The same idea applies to storage and vector infrastructure. Multiple agents can share the same foundational systems while remaining logically separated in terms of purpose, tools, memory, and context.
This is where efficient infrastructure design becomes critical. Shared services reduce costs and improve utilization, but logical separation remains essential to prevent cross-contamination between agents. The goal is to get both operational efficiency and contextual isolation.
That pattern makes it possible to scale an ecosystem of specialized agents without scaling infrastructure cost in a strictly linear way. In other words, organizations can support many specialized use cases without rebuilding the stack from scratch every time.
Why the Data Pipeline Still Matters
Even with all the discussion around models and algorithms, the real power of AI still comes from data.
Enterprise data is rarely neatly organized in one place, ready for use. It’s scattered across on-prem systems, cloud platforms, edge environments, data warehouses, object stores, and operational applications. It spans structured data, unstructured data, and everything in between. Any part of it might prove valuable for an AI use case.
The real challenge is access.
Most organizations are constrained by fragmented systems and siloed architectures that make it hard to ingest, transform, and unify their data. A modern AI pipeline needs to bring that information together into a scalable enterprise data lake or equivalent governed foundation.
That foundation then becomes the system of record enriched with metadata, lineage, and governance. That way, the data isn’t merely stored, but is discoverable and AI-ready.
Once that is in place, the same governed data can be used to generate curated datasets and multimodal vector embeddings. Those then power use cases such as retrieval-augmented generation, semantic search, intelligent assistants, model training, inference, analytics, and traditional BI workloads.
The important shift is that one governed pipeline can support many downstream uses. But that only works when access control and governance are embedded end-to-end. Trust matters just as much as performance.
Why Storage Is at the Center of AI Architectures
When people look at modern AI architectures, they often focus first on models, agents, or the inference layer. But at scale, the storage layer is often the most critical component.
Agentic AI is compute-intensive and deeply stateful. Agents reason over context, retain memory, retrieve history, and act over time. All of that depends on persistent, high-performance, reliable storage.
This is where questions like “where does an agent store its memory?” become central rather than theoretical. How does an agent retrieve historical information? How does it improve based on past outcomes? How does it maintain shared, yet isolated, context in a large system?
Storage becomes the system of record for agent state. It stores raw data, vector embeddings, key-value caches, short-term context, and long-term memory. In that sense, it becomes the durable layer that makes reasoning repeatable, explainable, and persistent over time.
Without that, agents become effectively stateless. Local memory may work in limited cases, but it quickly breaks down when systems need to scale, share context, or operate across multiple services.
That is why storage performance matters as much as capacity. In agentic systems, storage is no longer a passive backend tier. It sits directly on the critical path for reasoning, retrieval, and retrieval-augmented generation. At scale, storage performance becomes inference performance.
Governance, Durability, and Trust
The other reason the storage and platform layers matter so much is governance.
A shared data platform enables consistent governance, access controls, and historical traceability across both raw and transformed data. That is essential if organizations want AI systems that are explainable, auditable, and safe to scale.
As agent adoption grows, the amount of stored context grows even faster. It’s not just raw data expanding. Embeddings, metadata, and long-lived memory accumulate continuously. That means the storage layer has to support mixed workloads, vector and metadata access, scalable growth, and resilience by default.
At that point, the limiting factor is no longer just the number of models you can deploy. The question becomes how well your infrastructure stores, retrieves, governs, and manages memory, context, and state over time.
Storage is the foundation that enables reasoning, trust, and long-term value.
What a Modern AI Platform Needs
A modern AI platform requires more than storage or compute alone. It needs a set of capabilities that work together.
It needs to support querying structured business data. It needs distributed data platform capabilities, metadata, and governance. It needs scalable storage. It needs data preparation for both structured and unstructured workloads. And increasingly, it needs to operate across hybrid environments rather than in one isolated location.
The broader point is not that every organization must use the exact same stack. It’s that the platform remains open. AI workloads will continue to evolve, and new use cases and technologies will emerge. Closed architectures become constraints very quickly.
What teams need is a platform that consistently supports the full lifecycle: ingesting data, governing it, transforming it, preparing it, and making it available for AI workloads, in ways that can evolve over time.
Final Takeaways
As AI progresses from generative systems to agentic systems and beyond, the data layer becomes increasingly central, not less so. Context, memory, retrieval, state, governance, and performance all become core requirements. That is why data platforms must evolve into active systems that support the full lifecycle of AI, not just static repositories.
The second takeaway is that scale depends on shared infrastructure as much as on models. The organizations that succeed will be those that can support many specialized systems on a common, efficient, well-governed foundation.
And the third is that storage sits at the center of it all. It’s the layer that makes memory durable, retrieval fast, reasoning repeatable, and trust possible at scale.
Modern AI is only as capable as the data architecture beneath it. The real challenge is not simply training larger models or deploying more agents but building an open, governed, scalable data platform that can support continuous discovery, curation, retrieval, training, inference, and long-lived memory as AI systems evolve.
That is what turns data from a passive asset into an active foundation for AI.
Want the full session? Watch Himabindu Tummala’s talk on demand at the AI-Ready Data Summit.
