Sovereign AI: Who Holds the Keys?

Imagine: 2:14 a.m., a refinery control room. A pressure anomaly is spreading across three units. The shift lead turns to the AI assistant that has read every maintenance log and incident report from the last decade. It replies with a polite message: service unavailable for your organization.

Fiction? Only just.

In July 2025, after a new round of EU sanctions, Microsoft suspended Outlook, Teams and other core tools at Nayara Energy, operator of one of India's largest refineries, even though the company held fully paid licenses. Employees moved to a domestic email provider. Access returned days later, just before a Delhi High Court hearing.

That was email. Now picture the AI that reads your contracts, prices your bids and watches your plant.

That is the question Sovereign AI answers: who holds the switch?

What is Sovereign AI?

Sovereign AI is an organization’s ability to decide where its AI runs, what it learns from, what data it can access, which model it uses, who can see it, and who can turn it off.

In other words, sovereignty is about control across the AI stack.

It isn’t a flag on a data centre, and it isn’t simply a "made in India" label. It also isn’t the open-versus-closed debate. It is control across four layers.

Sovereignty isn’t owning everything. It’s never being forced to accept a decision you didn’t make.

Why does sovereignty matter for mission-critical industries?

In most businesses, an AI outage is an inconvenience. In mission-critical industries it is a safety incident, a compliance breach or a lost contract.

These industrial organizations possess something increasingly valuable: proprietary data and institutional knowledge. Engineering drawings. Source code. Product architectures. Manufacturing recipes. Maintenance history. Operational telemetry. Safety cases. Customer information. Supply-chain intelligence.

This information is often the competitive advantage itself.

Putting that data into an AI system therefore isn’t simply a technology decision. It is a decision about IP, security, regulatory exposure, business continuity and competitive advantage.

Nations already get it. The IndiaAI Mission has backed 20 indigenous model proposals: 12 large and 8 small language models. Enterprises running critical operations need the same instinct.

When does an AI system need to be air-gapped?

"Cloud-first" is a sound default, until your data is a blueprint, a grid topology or a defence tender. For that class of data, the safest network connection is none.

An air-gapped deployment removes three risks at once: data leaving the perimeter, dependence on outside connectivity, and silent changes pushed from outside.

NIST defines an air gap as an interface where systems are not physically connected and logical connections are not automated, with data transfer occurring manually under human control.

Our rule: match the gap to the classification. Not every workload needs isolation. The ones that do should never be forced online.

Open weights vs frontier models: the wrong debate?

Security teams learn this the hard way: a rented model may refuse to analyse real attack logs because they look like an attack. A self-hosted model works under your policies, not someone else’s.

The winning pattern isn’t picking a side. Frontier models for complex reasoning on non-sensitive work. Open weights for sensitive, regulated or offline workloads. Routed by policy, not by habit.

Bring your own LLM: don’t let your AI strategy expire

If replacing your model means rebuilding your platform, you don’t have an AI strategy. You have a dependency.

One of the biggest architectural mistakes an enterprise can make is allowing its AI platform to become permanently coupled to one model provider. Models will change. Prices will change. Capabilities will change. Regulations will change. Your business requirements will change.

Your AI strategy should not expire with your LLM contract.

Bring Your Own LLM (BYO-LLM) means the application layer is model-agnostic. You plug in the model you trust, whether frontier, open-weight or one of India’s emerging sovereign models, and swap it when a better, cheaper or safer option appears.

• Negotiating power stays with you.
• Data agreements stay with you.
• Model upgrades become a configuration change, not a project.

Choosing the right model: five questions before you sign

The most overlooked Sovereign AI feature: the kill switch

Here’s a simple test I believe every enterprise AI architecture should pass: can you stop it?

If an AI agent is connected to enterprise systems, it may eventually have the ability to:

• Read sensitive information
• Generate engineering artifacts
• Modify software
• Trigger workflows
• Invoke APIs
• Make recommendations
• Potentially execute actions

At that point, the AI is no longer merely answering questions. It is operating inside your enterprise. A sovereign architecture therefore needs controlled mechanisms to stop, isolate, revoke, roll back and recover.

The kill switch should not be an emergency button invented after deployment. It should be an architectural capability.

Sovereignty means shrinking the left column and owning the right one. Design so that when the AI goes off, the business doesn’t. Human-in-the-loop isn’t the brake. It’s the steering wheel.

The sovereignty test

Five statements. Be honest about how things run today.

The last word

The next decade of AI in critical industries won’t be won by whoever has the smartest model. Models will keep getting smarter, cheaper and more interchangeable. It will be won by the organizations that control the stack those models run on.

Intelligence is becoming a commodity. Control is not.

So before your next AI decision, ask the only question that matters at 2:14 a.m.: who holds the switch?

At tiramai, we build AI-native platforms for industries that can’t go dark: sovereign by design, model-agnostic by default. Let’s talk.

Sources

• PTI, Microsoft suspends services to Nayara, firm sues in Delhi High Court, 28 July 2025
• PTI, Microsoft restores services to Nayara ahead of HC hearing, 30 July 2025
• Outlook Business, Microsoft suspends Nayara’s digital tool access: developments so far
• CXO Digital Pulse, IndiaAI Mission backs 20 sovereign AI models, July 2026
• NIST Computer Security Resource Center, air gap; NIST, AI Risk Management Framework
• OECD, AI Principles
• tiramai, Security and Responsible AI architecture; Responsible AI, through the TRUST AI framework