For over 15 years now, NVIDIA GTC has been where the big shifts in AI infrastructure become legible. This year, several threads are converging at once: sovereign AI, agentic systems, physical AI, and the rise of AI factories as an operating model.
Underneath the variety is a common pressure. AI is moving from generating content to driving work. That changes what enterprises have to build for: real-time access to operational data, clear controls over what systems can do, and infrastructure that can hold up under unpredictable demand.
In that environment, the difference between a promising pilot and a dependable system is not whether a model can produce a convincing answer. It is whether the data platform can expose the right data to the right process at the right time, with guardrails that actually hold.
This is where sovereignty becomes operational. True data and AI sovereignty means you can access, manage, and control your data, wherever, whenever, and however you want to use it, securely and compliantly, across clouds and your premises.
Said plainly: you need control without handcuffs. You need governance without making the system unusable. And you need performance that doesn’t collapse the moment workloads become agentic.
That’s the enterprise architecture conversation showing up at NVIDIA GTC 2026.
Why this conversation is converging now
NVIDIA has elevated sovereign AI from a broad idea to a dedicated set of sessions, framing it as a systems challenge tied to AI factories, national and regional requirements, and real deployment constraints.
Across the other major themes as well, the pattern is similar: capabilities are advancing quickly, and the practical questions are catching up. When autonomy increases and AI starts touching more of the operating model, the gap between “we can build it” and “we can run it safely” gets much harder to ignore.
The market expectations around agentic AI are also becoming more specific. Gartner predicts that by 2028, 15% of day-to-day work decisions will be made autonomously through agentic AI, and 33% of enterprise software applications will include agentic AI.
But Gartner’s warning is the important counterweight: it also predicts over 40% of agentic AI projects will be canceled by the end of 2027, citing issues like cost, unclear business value, and insufficient risk controls.
Taken together, that’s a signal that the next phase of agentic AI will be decided less by experimentation and more by architecture and what actually holds up in production. The teams that succeed will be the ones that treat sovereignty as an operating requirement and make the platform decisions that let autonomy run under control.
The three data platform decisions that matter most
When teams say they are building sovereign AI, the work tends to collapse into a few high-impact decisions. They are not flashy, but they determine whether agents can move past pilots.
1. What counts as usable operational data for agents
Agents are only as effective as the data they can safely touch. In practice, that means moving beyond static snapshots and curated knowledge bases toward governed access to systems of record, with clarity on what is in scope and what “fresh” means for the use case.
This is where programs quietly narrow to what is easy instead of what is valuable. A common first step is a support copilot that drafts responses based on documentation. The harder, higher-impact step is when the system can verify entitlement, check order status, reference policy, and initiate the next action without creating new risk.
That jump is not a “model upgrade;” it is a data platform decision.
Getting this right is about usability and control at the same time. Teams need to decide which operational sources are usable for agentic workflows, how data quality is validated, and what constraints apply when the output triggers a downstream action.
2. Whether control and auditability reach the runtime
Sovereignty is not just where data sits. It is whether you can enforce how data is accessed and used.
For agentic systems, that means permissions and policy cannot live in documentation. They have to be enforced at runtime: role-based access, purpose-based constraints, data minimization, and guardrails on what tools can be called. And when something goes wrong, you need the ability to trace what was accessed, what action was taken, what policy was in effect, and why.
This is also what makes autonomy defensible. If an agent is allowed to trigger a workflow, you need a clear answer to basic questions: what data did it retrieve, what did it change, and under what controls. Otherwise, every incident becomes a trust event, and every expansion becomes a debate.
3. Whether hybrid operability is designed in
A lot of sovereignty requirements exist because the real world is hybrid: data lives across clouds and premises, and “move it all somewhere else” is not an answer most enterprises can accept.
So the platform decision is whether governance can travel. Can you keep consistent control and auditability even when workloads and datasets span environments? Can you keep the system usable without trading away operational boundaries?
In practical terms, hybrid operability means the governance model does not change just because the workload moved. The same access boundaries apply, and the same audit story holds, regardless of where the data lives or where the agent runs.
Where to find EDB at GTC
If you want to pressure-test these decisions against real architectures, NVIDIA GTC is a good place to do it. It’s one of the few forums where the conversation spans the full stack: infrastructure, data platforms, and the realities of running AI systems under control.
EDB will be there in San Jose for GTC (March 16-19) at Booth 3414 to talk through these platform decisions in concrete terms, including live demos of Warehouse Analytics and the Sovereign Data and AI Factory. You can book your meeting with our team now.
EDB CTO Quais Taraki will also be speaking on March 16 in the Expo Theater. Add his session to your schedule.
And if you want a deeper blueprint for the platform conversation, EDB’s O’Reilly business guide, Building a Data and AI Platform with PostgreSQL, is part of the on-site experience at GTC. Download your free copy here.