VAST Data Co-Founder, Jeff Denworth, on building the AI grid, a distributed, interconnected AI infrastructure network.
The AI grid is a geographically distributed, interconnected AI infrastructure fabric that extends intelligence across metro, edge, and national networks, engineered to put compute where latency and sovereignty requirements demand it. Nscale is helping build it.
Compute is the part of the grid everyone counts. Whether the grid actually works depends just as much on the data those GPUs need, and on the infrastructure required to make that data available at every site the grid reaches.
Building it requires a global ecosystem of partners. Among them is VAST Data, which has spent a decade on what it describes as the AI operating system: a unified platform that brings storage, database, and compute together to provide the data foundation for modern AI and agentic workloads.
Working with leading organizations across government, global telecommunications, finance, and healthcare, Denworth observes the same structural failure repeat. The compute catches up. The data strategy doesn't.
We sat down with Denworth to talk through the questions that distributed AI is forcing organizations to answer. What breaks when you deploy AI inference across hundreds of edge sites without a unified data layer? How do you build sovereign AI infrastructure without creating silos? And what does an organization actually need to understand about data architecture to survive the coming explosion in agentic AI?
Denworth on why the data layer decides whether distributed AI infrastructure delivers. Part of Nscale's AI grid series.
This conversation has been edited for length and clarity.
Why is data the real bottleneck when building the AI grid?
You can't have AI without data. We can try to control where AI workloads run to build the best possible systems, but that's not how AI is used in practice. Users travel all around the world. Organizations partner and collaborate across the globe. So being able to get an experience as close as possible to those users, so that AI can be responsive and productive, means you have to think seriously about how your data gets to the machines you're distributing all across the grid.
That creates real problems with respect to data management and governance. It also creates excessive infrastructure utilization, because you have to copy data across many different locations and ensure those copies stay consistent.
The answer is a unified data layer that can make data available wherever compute is running, without forcing you to create and manage copies across every location. That gives you the experience of local data at the edge, while maintaining consistency and control across the grid. We solve these problems through the VAST AI Operating System. Within it, VAST DataSpace federates distributed data and computing environments so applications can access data across edge, data centre and cloud locations without creating another set of isolated data silos.
What breaks first without a unified data layer when you're deploying hundreds of sites across a telco infrastructure?
Pretty much everything. The first question is: whose problem is it to reconcile getting the right data to the right location at the time it's being requested?
You could solve that with an unlimited budget by making copies across your entire estate, but that's expensive and wasteful. You could try to solve it with a large operations team, but that means people constantly reconciling permutations and transformations of data happening across the edge, while AI tools are generating new data at every site.
These are complex problems that legacy technologies were never designed to address. Within the VAST AI Operating System, the VAST DataSpace is designed to remove that reconciliation burden: it spans tens to hundreds of sites with real-time, consistent data and a single management plane, rather than leaving you to manage hundreds of separate systems. The goal is to give each location the experience of having the data it needs locally, without having to create and manage a separate copy at every site. You don't have to worry about where it is. It's just where it needs to be, when you need it, as fast as it can possibly be.
How does VAST enable sovereign AI across distributed infrastructure without creating data silos?
The challenge is building unified infrastructure that doesn't require you to build different systems for different governments or customers, while still providing the security controls that allow for data isolation across organizational or geographic boundaries when that's required.
We've done substantial work with some of the world's largest national security agencies, and you can imagine they have strict guidelines around who can see what data. Those same classification principles can be embedded into a system as a set of automated policies and applied across a distributed data center footprint, ensuring certain data doesn't cross certain barriers or boundaries.
That governance is all part of the VAST AI operating system. We're extending it further with what we call the Policy Engine, an inline policy enforcement capability designed to govern the information flowing between agents, tools and data sources as agentic systems execute.
Agents running through VAST AgentEngine can be observed and governed against those policies. Here imagine deploying agents natively from the VAST OS, with a set of explicit criteria that allow or disallow specific actions. The Policy Engine sits in-band between an agent and all of its data: memories, tools, and RAG-contextualised datasets. Whether you're using explicit rules or fuzzy rules interpreted by large language models to de-identify data, policies can be continuously enforced as agents interact with data and tools.
When a telco moves from connectivity provider to AI service provider, what has to change about how it thinks about data?
AI is a lot more than just compute, networking, and storage. That is why the infrastructure layer increasingly needs to behave like an operating system rather than a collection of point products.
If you look at the reference architectures that service providers use to deploy infrastructure, they're actually quite simplistic. The question then becomes whether that meets the needs of customers who are going to be AI infrastructure consumers from a telco. The reality is that the activities involved in building AI pipelines don't fit neatly into the traditional categories of compute, networking, and storage.
Data strategy is critical, particularly as you try to monetize a network with distributed data centers. You have to unify everything without exploding your costs or management overhead.
But the more interesting consideration is what sits behind that compute. Pre-training requires vast amounts of data. Model training requires data fed into GPUs and additional data output from those GPUs. You need to organize all of that with file systems and ETL tools. Once a model is trained, whether built in-house or deployed off the shelf, there's a whole data strategy required for auditing and logging inference activities. Inference engines need access to the broadest possible data sets: enterprise file systems, SaaS services, enterprise data warehouses. And everything an agent or inference operation does needs to be captured in tamper-proof, auditable spaces, so you can conform to requirements like the EU AI Act.
For telcos, having a comprehensive strategy that brings together data services and computing services is what separates an AI infrastructure provider from a GPU reseller.
As AI becomes more affordable and accessible, what changes about the storage and data architecture enterprises need to understand?
Agentic AI is unlocking enormous levels of efficiency across organizations of every type, and we're still in the earliest days. Today's projects are relatively small and exploratory. But move one to two years forward and we'll see an explosion of agency: agents responsible for tasks, agents managing other agents, layers of automation building on each other.
When you get to that level of agency, every organization will need to function like a supercomputing organization, because it will require a tremendous amount of machinery to power that level of automation. And those agents are constantly accessing and utilizing data to accomplish their tasks, while generating enormous amounts of exhaust that has to be captured, monitored, and governed for auditability and explainability.
Here's the core problem: all these new applications need to access data that has been around for 10, 20, or 30 years, sitting in systems and architectures conceived of decades ago. None of that is going to be appropriate for what's coming. That's why VAST Data was created. If deep learning and agents were going to become the dominant way businesses operated, every single layer of the stack needed to change. We've built a fully parallel systems architecture, so enterprises can support increasingly data-intensive AI and agentic workloads without continually adding new infrastructure silos.
Based on your work with enterprises in finance and healthcare, what are the major challenges they face deploying AI at scale?
We're still in such early days that many of these organizations haven't yet achieved scale. But even the early work is showing that enterprises need to fundamentally rethink how they operate as they move into AI and agency.
From an infrastructure perspective, you need to meet the demands of a new computational model. At the same time, organizations have to bridge legacy systems and applications into what they’re building with AI and agency, while maintaining the security and safety of their existing stack.
We've distilled down the reason AI hasn't yet taken off in every enterprise to one thing: trust. If an organization can't trust what it's doing, it will fear moving forward as fast as it should. Have you fine-tuned the right data into your models? Once your models and agents are deployed, do you understand what they're doing? Do you understand the data they're accessing and what they've been trained on? And can you roll forward or roll back with confidence?
That is why trust, governance and observability are built into the architecture of the VAST AI Operating System, with activity recorded in auditable, tamper-proof logs that support regulatory requirements, while providing genuine observability into what's happening across your AI environment. We've been working on this architecture for 10 years. When the generative AI market took off, we had the right product at the right time for enterprises ready to step into this world. And we're going to keep building as customers challenge us to make AI even easier and more safe.
.png)
