95% Say Sovereign AI Matters. Only 29% Are Acting on It.

NTT DATA released its 2026 Global AI Report on May 14. The subtitle gives away the theme: A Playbook for Private and Sovereign AI. It draws on two studies covering nearly 5,000 senior decision-makers, more than a dozen industries, over 30 markets, and five regions. The finding underneath all of it is that enterprise AI is outgrowing the infrastructure it runs on, and the pressure point is data.

For years the job of enterprise architecture was to move data fast, across systems, clouds, and borders. AI assumed that same freedom and built on it. Now privacy rules, regulators, and national data laws pull the other way. Sensitive data has to stay in defined places. Models have to run under tighter control. The report's word for this is blunt: jurisdiction has become a core architectural constraint.

Private AI and sovereign AI are not the same thing

The two terms get used as if they mean one thing. They don't, and the difference decides where you can run AI.

Private AI is about protecting your data. It keeps sensitive enterprise information controlled, limits who and what can reach it, and reduces exposure. Sovereign AI is about jurisdiction. It makes sure the AI system, the data it touches, and the environment it runs in all meet the rules of a place that has authority over you: a country, a region, or a regulator.

You can have one without the other, and regulated institutions usually need both. A model can be private, sealed inside a vendor's cloud tenant, and still sit under a jurisdiction that can compel the data or a law you can't satisfy. We walked through that exact trap for law firms in private LLMs and data sovereignty. Privacy asks who can see the data. Sovereignty asks whose laws govern it and where it physically lives.

Why jurisdiction became an architecture problem

Data can still move. It just can't always move the way AI wants it to.

Modern AI, and retrieval and agents in particular, expects continuous access to data wherever it sits. The usual pattern pulls data toward a central model. Jurisdiction cuts across that. Some data can't cross a border. Some can't land in a given cloud region. Some can't sit on hardware a foreign government could reach. So jurisdiction now decides three things at once: where the data lives, where the model runs, and how the whole system is governed.

Nearly 60% of the AI leaders NTT DATA surveyed called cross-border data restrictions a major challenge. When the rules say data can't leave a region, that decision lands in the architecture, because the model has to come to the data instead of the other way around. Systems built for centralized, borderless flow hit a wall the moment you bolt those rules onto them. The report's first finding puts it plainly: the constraint is no longer the model. It's compute, data access, security, and locality.

Everyone sees it. Few are moving.

The gap should worry anyone running AI in a regulated setting. More than 95% of organizations told NTT DATA that private and sovereign AI are important. Only 29% are prioritizing sovereign AI in a concrete, near-term way. The understanding is nearly universal. The action is roughly a third of it.

The report's read is that this gap is turning into a divide. The organizations redesigning early, aligning infrastructure, governance, and operating models up front, are moving from pilots to scaled deployments faster than the ones layering AI onto architectures that were never built for control or locality. The leaders aren't waiting for the rules to settle. They're building for where the rules are clearly heading.

The rules are moving too

The regulatory picture is the other half of the story, and it's unsettled enough that the safest move is to design for control rather than bet on a date.

In Europe, the AI Act's obligations for high-risk systems were set to apply from August 2, 2026. A proposed package known as the Digital Omnibus would push the high-risk deadlines back, to December 2027 for standalone high-risk systems and to August 2028 for AI built into regulated products, but as of mid-2026 that deferral still has to be formally adopted, and the transparency obligations under Article 50 stay on the earlier schedule. The dates are in motion. The direction of travel, toward documented governance and data control, is not. You can read the current state in Holland & Knight's analysis and Gibson Dunn's summary of the omnibus.

The United States has no single federal AI law, but a state patchwork is forming under it. The Texas Responsible AI Governance Act took effect January 1, 2026, enforced by the state attorney general, with a safe harbor for firms that substantially follow the NIST AI Risk Management Framework. Colorado passed the first broad state AI law, then repealed and replaced it in May 2026 with a narrower automated-decision statute that starts in 2027. The shape of US AI regulation is still being drawn.

None of that is the whole reason sovereignty matters here, though. The sectoral rules already on the books act like sovereignty rules without using the word. HIPAA, the Gramm-Leach-Bliley Act, and CMMC and ITAR for defense work all turn on the same question: where does regulated data go, and who can reach it. Sovereign AI is the architecture that gives a defensible answer.

What sovereign-ready architecture looks like

Strip away the label and sovereign AI comes down to controlling three layers: the infrastructure the model runs on, the data it touches, and the model itself. Control all three and the word stops being a slogan. In practice that means a handful of concrete things:

• Run the model where you choose. The compute that serves the model sits in an environment you control, your data center, your private cloud, or an air-gapped enclave. That's how you pin the jurisdiction in place.
• Keep the data put. Documents, prompts, and outputs stay inside your boundary and your region, so nothing crosses a border you didn't approve.
• Enforce permissions inside retrieval. The system checks who is allowed to see a document before it feeds it to the model, so a search can't surface something the person asking isn't cleared for.
• Log every interaction locally. Each prompt, source, and answer is recorded where you can produce it for an examiner, without filing a request with a vendor for its logs.
• Give agents a scoped identity. Any agent acts within set permissions, and every action traces back to an owner.

None of this requires one specific product. It does require a deployment model that keeps the data and the model on your side of the line, because you can't enforce locality on infrastructure you don't control.

Where Cognetryx fits

We build private AI for regulated institutions, and the NTT DATA report describes the wall we set out to get around. Keep the model and the regulated data in your own environment, and the privacy question and the sovereignty question get answered by the same decision.

• Your data stays in your boundary. Prompts, documents, and outputs don't transit a shared public AI service, so a provider's region or a cross-border rule isn't a problem you inherit.
• Retrieval respects permissions. The retrieval layer enforces who can see what before anything reaches the model. More on permission-aware RAG.
• Agents run with scoped identity. Agent actions are bounded, logged, and traceable to an owner. More on on-premises agents.
• Answers carry citations. Every output traces to its source, which is often the difference between an AI an examiner will accept and one it won't.

If you're weighing where AI should run, how to choose between cloud and on-premises AI governance walks through the trade-offs against data sensitivity, regulators, and audit needs.

Sovereign AI is mostly an architecture decision wearing a policy label. The 29% who are already building for it made the call early, while the rules were still moving. NTT DATA's own read is that the distance between them and everyone else is set to widen.

Keith Kennedy, CISSP

Founder, Cognetryx

Keith is an IT thought leader with nearly 20 years of experience architecting secure technology solutions for regulated industries. He holds a CISSP certification and has advised enterprise companies on HIPAA, SEC/FINRA, and GDPR compliance.