Sovereign AI for Government: Data, Operations and Air-Gapped AI

Watch the full session · 56 min

Nations want to produce AI, not just consume it. In a Data Science Connect webinar, leaders from NVIDIA and HPE defined sovereign AI, explained data, operational and technology sovereignty, and described why sensitive-data rules and air-gapped environments keep many AI projects from reaching production.

Governments are moving from consuming AI to producing it, and that shift puts data sovereignty at the center of public sector AI strategy, according to speakers at a Data Science Connect webinar brought to the audience by Hewlett Packard Enterprise and NVIDIA. The panel featured Shilpa Kolhatkar, Global Head of AI Nations at NVIDIA; Bhuvaneshwari Guddad, Chief Technologist, WW A&PS AI Services, at HPE; and Steve Heibein, Federal AI Chief Technologist at HPE. Brian Mink, co-founder and president of Data Science Connect, hosted.

The speakers defined sovereign AI, laid out the principles behind it and described what happens when sensitive data can't leave a room, let alone a country.

What is sovereign AI?

Kolhatkar, who works with governments around the world to turn AI strategies into execution, said everyone is rapidly adopting AI, so "the critical question is now who is producing it?" Nations, she said, want to become creators of AI, not just consumers.

She defined sovereign AI as "a nation state's capability to develop and deploy AI using its own infrastructure, its own local data sets," along with a skilled workforce and business ecosystem. A nation's data is akin to a natural resource, she said, and sovereign AI reflects a need for self-reliance and less vulnerability to foreign infrastructure and models.

The recipe nations follow ties AI investment to policy outcomes and builds on three ingredients: a domestic ecosystem of investors, developers and entrepreneurs; local AI skills, including upskilling the current workforce; and foundation models built on local data and hosted on local infrastructure. Some countries are building sovereign clouds through national telecom providers, regional cloud providers or energy companies, she said. Indonesia, for example, partnered with its national telecom operator.

Sovereignty isn't limited to the public sector, Kolhatkar said. Financial services, healthcare, pharmaceuticals and critical infrastructure such as energy and transport all handle data that must be protected, and every region has its own evolving rules.

What are the principles of digital sovereignty?

Guddad grouped sovereignty into three areas:

• Data sovereignty: data stored and processed within national borders, privacy and protection so people know how their data is used, local governance under regional laws, and representative training data to avoid unconscious bias.
• Operational sovereignty: full visibility and control over IT operations, with frequent audits, transparency and monitoring under regional law.
• Technology sovereignty: control of the full stack, from infrastructure to models and algorithms, with interoperability to avoid vendor lock-in.

Why do sovereign AI projects stall before production?

Storing data locally is only the start, Guddad said. Once data must stay in-country, model development, training, deployment and governance must happen there too. So must monitoring and incident management. Many hyperscalers store data locally, she said, but customers ask whether auditing and monitoring are still controlled by foreign entities. When that requirement can't be met, applications don't move to production.

She described one customer so concerned about leaks that HPE built an entire AI application on synthetic data. To test accuracy against real data, the customer's data scientist took a laptop into an isolated room with no access to other devices. "If we are doing this way then how are we going to scale?" she said.

Some federal customers also distrust open-source models hosted in the cloud, she said, quoting one: "It's a open source model so I cannot trust." For them, models, data and inference should all be on premises.

How do you deploy AI in air-gapped environments?

Guddad described two operating models. In sovereign management, the platform is operated by partners in the same geography without crossing national borders. In disconnected management, the environment is air-gapped. Teams have to work with IT to decide which firewall ports can be opened, staff must work on site because there is no remote access, and customers may scan every Docker image to understand each component.

In Q&A, she said the most basic challenge is data. "If you don't have a right data you cannot have the right accuracy," she said. Skills to maintain systems over the long term and integration with existing applications follow close behind.

What does an on-premises sovereign AI platform include?

Heibein described HPE Private Cloud AI, co-engineered with NVIDIA, as a turnkey generative AI platform delivered in a rack and usually running within 24 hours. It includes NVIDIA AI Enterprise software, HPE AI Essentials as the interface, a three-step retrieval-augmented generation setup and HPE's private cloud control plane for provisioning, Kubernetes, cluster management and observability. Because storage, data, networks and security stay on premises, he called it a turnkey route to sovereign AI.

Asked whether it supports only NVIDIA models, Heibein said, "it's wide open." Teams can import models from Hugging Face, GitHub and elsewhere. On cost, Guddad said HPE GreenLake is pay-as-you-use with support bundled. Heibein said that over five years the total cost of ownership can be as little as about half that of commercial clouds.

Kolhatkar said the countries making progress share a focus on sovereign infrastructure, data infrastructure, sovereign large language models, talent and startups, with early public-private partnerships and open models.

More AI governance discussions in Data Science Connect webinars

Data Science Connect's 2027 schedule includes 24 free live webinars across four series: the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. Topics include data governance, security and deploying AI in regulated environments. See the webinar schedule.

Watch: Sovereign AI: Keeping Your Data Secure and Controlled on YouTube
Event: Webinars · More coverage: Webinars insights