A couple of weeks ago, I posted a LinkedIn poll with a simple question:
What does “sovereign AI” mean?
Does it mean running AI on infrastructure you own? Or running it on infrastructure located in your country or region?
The results were almost exactly 50–50.
That surprised me. Not because I expected everyone to agree. “Sovereign AI” is a relatively new term, so different interpretations are understandable.
What surprised me was who disagreed.
Most respondents were technical people. They were interested enough in the subject to answer a LinkedIn poll about it, which is already a high bar. They also weren’t concentrated in one group. Engineers were split. Infrastructure specialists were split. CEOs, founders, and vendors were split.
There wasn’t a job title that indicated who understood sovereign AI “correctly.” People simply meant different things by it.
That’s a problem.
The words matter
Technology discussions often treat disagreements about terminology as pedantic. Who cares what we call it if everyone knows what we mean?
Except we often don’t.
Take the two definitions from my poll:
If sovereign AI means running AI on infrastructure you own and control, that creates one set of requirements.
If it means running AI on infrastructure located within your country or region, that creates another.
The two can overlap, but they don’t have to.
You can run workloads on infrastructure in your country without owning it. You can own the infrastructure but operate it somewhere else. You can control the infrastructure while depending on someone else for the model. You can control the model but run it on somebody else’s infrastructure.
These differences affect procurement, architecture, regulation, security, cost, and the outcome you’re trying to achieve.
When two people say they want sovereign AI, the next questions should be:
Sovereign from whom? And in what sense?
Then there’s “local AI”
I ran into a similar problem when someone described a coding session as “local AI.”
I assumed the model was running on their machine or infrastructure. It wasn’t.
They were using a model through the Kilo Gateway. The coding agent was running on their machine rather than in a cloud-based agent environment.
That’s a reasonable setup, but it shows why the term is ambiguous.
When someone says “local AI,” what exactly is local?
The model?
The inference?
The agent?
The data?
The development environment?
The network connection?
If I’m evaluating a system against a requirement such as “our data cannot leave this machine,” those distinctions matter.
AI has a vocabulary problem
We’re introducing terms faster than we’re agreeing on what they mean:
Sovereign AI. Local AI. Private AI. Self-hosted AI. On-prem AI. Open models. Open-weight models. Cloud agents.
These terms can be useful shorthand. They can also be marketing language. Increasingly, they’re both.
The risk is that we make decisions based on a label instead of the underlying requirement.
One company might say it needs sovereign AI because it doesn’t want sensitive data subject to another country’s jurisdiction. Another might want to avoid dependence on a particular cloud provider. A third might want to run open-weight models on hardware it controls. Some might mean all three.
Those are different problems. They may lead to the same architecture, but they may not.
Ask what problem you’re solving
We probably won’t create one definitive AI glossary that everyone follows. I’m not sure we should try.
Instead, we need to ask one more question when these terms come up:
What problem are you actually trying to solve?
If someone wants sovereign AI, ask whether they mean data, infrastructure, model, or jurisdictional sovereignty—or some combination.
If they want local AI, ask what needs to be local.
If they want an open model, ask what they mean by open.
If they want an AI agent running locally, ask whether they mean the agent, the model, or both.
These questions sound obvious. My LinkedIn poll suggests they aren’t being asked often enough.
We don’t need perfect definitions. We need shared understanding.
The poll didn’t show that half the respondents were wrong. There may not be one correct answer.
It showed that technically informed people use the same term to describe different things.
That matters as terms such as sovereign AI appear in architecture decisions, government policy, procurement documents, and vendor pitches. In those contexts, ambiguity becomes expensive.
The next time someone tells me they want sovereign AI, I’m going to ask a slightly annoying question:
What, exactly, needs to be sovereign?
Until we answer that, we may not be talking about the same thing.
In financial planning, I see the same pattern — clients say 'retirement corpus' or 'wealth creation' without defining the goal underneath.
The question 'what problem are you solving?' applies as much to money as to AI. Until we specify, the label is just a placeholder.