Sovereign AI Is Not Just About Where the Model Runs

Real AI sovereignty starts with the infrastructure.

There is a growing push toward sovereign AI. Countries want their data to stay within their borders. Governments and enterprises are deploying open-source models on local infrastructure. The reasoning makes sense: reduce dependency on a small number of global AI providers, gain more control over sensitive data, and build national AI capabilities.

But I think we need to ask a harder question: If you can run the model locally, does that really make your AI sovereign?

What Does Sovereign AI Really Mean?

Sovereign AI is commonly understood as the ability of a country or organization to control how AI is developed, deployed, and operated under its own governance requirements. That includes data, models, infrastructure, and the dependencies required to keep the system running.

Installing a model is increasingly the easy part. Running it reliably at scale is a completely different problem.

A model may work perfectly when one person asks a question. It may still work when ten people use it. But what happens at one hundred concurrent requests? Ten thousand? A million?

What happens when AI is no longer a chatbot used occasionally, but is embedded into thousands of business processes, applications, and autonomous operations running continuously?

That is where AI becomes an infrastructure problem.

Why AI Infrastructure Matters at Scale

The numbers give us some perspective on just how large that problem is.

OpenAI’s Stargate initiative was announced with a plan to invest $500 billion in AI infrastructure over four years. OpenAI later said its infrastructure planning had moved beyond the original 10 GW target .

Anthropic announced a $50 billion investment in American computing infrastructure in 2025. In 2026, it committed more than $100 billion over ten years to AWS technologies and capacity . Anthropic says it already uses more than one million AWS Trainium2 chips, while its expanded Google Cloud relationship includes plans for access to as many as one million TPUs .

These are announced plans and commitments across different timeframes, rather than a single measure of infrastructure already deployed. They nevertheless show the scale of investment behind frontier AI.

These are extraordinary numbers, and they tell us something important. The competitive advantage of today’s leading AI companies is not only their models. It is also their ability to train, serve, and scale those models across an enormous infrastructure footprint.

Data Residency Does Not Remove Infrastructure Dependencies

This creates another dimension to the sovereignty discussion. Imagine a country takes an open-source model, deploys it in its own data center, and keeps all of its data inside its borders. From a data residency perspective, it has achieved something important.

But the GPUs may still come from NVIDIA or another foreign supplier. The memory comes from a small number of global manufacturers. Storage, CPUs, networking equipment, and many other critical components come through complex international supply chains.

Even electricity generation, cooling, and data center construction become part of the equation once AI infrastructure reaches sufficient scale.

The model may be local. The data may be local. But much of the technology required to operate that model may not be.

That does not mean countries should stop investing in sovereign AI. Quite the opposite. It means we may need a broader definition of what sovereign AI actually means.

Can Useful AI Require Dramatically Less Compute?

Perhaps the bigger R&D opportunity is not simply recreating what OpenAI, Anthropic, or Google have already built.

If the current generation of AI requires hundreds of billions of dollars in infrastructure to operate at global scale, should every country try to reproduce the same architecture on a smaller scale? Or should part of the research agenda be focused on a different question?

How can we achieve useful AI with dramatically less compute?

• Models that require fewer accelerators
• Architectures designed around efficiency rather than unlimited compute
• Smaller specialized models for specific domains
• Better routing between models
• More efficient inference
• New approaches to memory, context, and reasoning
• Hardware and software designed together rather than independently

That could become a much more meaningful form of technological independence. Because otherwise there is a risk that we simply move the dependency.

We reduce our dependency on an American AI API, but increase our dependency on imported GPUs. We keep the data inside the country, but depend on foreign accelerators, memory, storage, and networking equipment to process it. We call the model sovereign, while the infrastructure underneath it remains global.

Should Every Enterprise Build Its Own AI Infrastructure?

There is a similar question at the enterprise level. For certain industries, governments, and highly regulated workloads, running models on private infrastructure can absolutely make sense. Data sensitivity, regulation, latency, or national security requirements may justify the investment.

But it does not automatically follow that every enterprise should build its own AI infrastructure. The economics of serving AI at scale are fundamentally different from installing software on a server.

Enterprises therefore need to separate three questions that are often mixed together: where their data resides, which models they use, and who owns and operates the compute infrastructure underneath those models. Those do not necessarily need to have the same answer.

AI Sovereignty Requires a View of the Entire Dependency Chain

The next phase of AI strategy, both for companies and countries, should go beyond choosing between proprietary and open-source models. We need to start thinking about the entire dependency chain.

Data sovereignty is important. Model sovereignty is important. But real AI sovereignty ultimately requires thinking about compute sovereignty as well.

And perhaps the most important innovation will not be building an even larger model. It may be figuring out how to need less infrastructure in the first place.

Frequently Asked Questions About Sovereign AI