Running a model locally is useful. Real sovereignty means retaining control over the model, data, policies, infrastructure, and, crucially, the exit door.
The word sovereign is having a very good year.
Every week, another product, platform, or press release promises “sovereign AI.” Usually, this means an open model running in a particular country, private cloud, or company-owned data center.
Those are meaningful advances. They solve real problems.
But a model does not become sovereign simply because you moved it into the basement and gave it a server rack.
We use the word sovereign at Backboard because it describes an important shift in enterprise computing. But I worry the term is being stretched until it means little more than “the deployment option we happen to sell.”
Sovereignty should mean something more demanding.
Sovereignty Is the Freedom to Decide
To me, AI sovereignty is an organization’s ability to make every meaningful decision about how its AI operates.
Where does the model run?
Which model do you use?
Can you modify or replace it?
Where does the data live?
Who controls the policies, memory, and access?
Can you move the workload without rebuilding the entire application?
That last question is the real test.
You can run a model entirely on infrastructure you control and still be deeply dependent on someone else.
The server may be yours, but the architecture can still belong to your vendor.
If your model cannot be adapted to your business, how sovereign are you?
If essential reasoning is locked behind a third-party API, how sovereign are you?
If your deployment options depend on a vendor’s roadmap, how sovereign are you?
If switching providers requires an 18-month migration, three consulting firms, and a small support group, you probably do not have sovereignty. You have a long-term relationship.
Real sovereignty is not isolation from every external provider. That is neither realistic nor especially desirable.
It is having meaningful control, credible alternatives, and the practical ability to leave.
Infrastructure Is Only One Layer
Most discussions about AI sovereignty begin and end with infrastructure.
Where is the data centre? Which cloud region is being used? Does the workload remain in-country?
These questions matter. But they cover only one layer of the system.
An enterprise AI application may rely on:
- A foundation model
- Post-training and model weights
- Organizational data
- Retrieval and memory systems
- Policies and permissions
- Application logic
- Inference infrastructure
- Deployment and monitoring tools
Controlling one layer does not automatically give you control over the others.
You can own the hardware while renting the intelligence. You can use an open model while remaining trapped in proprietary infrastructure. You can keep the data in Canada while sending critical parts of the workflow through services you cannot inspect, modify, or replace.
The residency map may look reassuring. The dependency map may tell a different story.
What Happens When the Model Learns Your Business?
Until recently, most enterprises treated foundation models as essentially fixed.
The company’s knowledge lived around the model:
- In system prompts
- In retrieval systems
- In vector databases
- In external memory
- In application code
- Occasionally, in a very large PDF called 'FINAL_v7_revised_ACTUAL_FINAL.pdf'
That architecture made sense when adapting a model was slow, expensive, and technically difficult.
But that assumption is changing.
As post-training becomes faster and more affordable, organizations can increasingly teach models their terminology, policies, workflows, institutional knowledge, and domain expertise.
The model stops being merely something the organization consumes. It becomes an asset the organization shapes and continually improves.
In some cases, it may become a meaningful part of the company’s intellectual property.
That changes the sovereignty question.
If years of institutional knowledge have been incorporated into a model, control over that model is no longer a minor procurement detail. The organization needs to know whether it can preserve, move, improve, audit, and continue operating that capability if a provider, price, licence, or regulation changes.
The smarter the model becomes about your business, the less comfortable you should be with someone else controlling the off switch.
Portability Is the Part Everyone Forgets
A sovereign architecture should let an organization choose between open and proprietary models.
It should allow models to be post-trained using the organization’s own knowledge.
It should allow them to be quantized and optimized for different hardware.
And it should allow the organization to decide where they execute.
Sometimes that will be a public cloud.
Sometimes it will be a private cloud.
Sometimes it will be infrastructure inside the organization’s own facilities.
Sometimes it will be a workstation, laptop, vehicle, factory, or edge device sitting directly in front of the person using it.
The important point is not that one of these options is inherently sovereign.
It is that the organization can choose among them, and change its mind later.
Models will change. Hardware will change. Providers will change. Prices will change. Regulations will change. Someone will rename a product, move three features into a new pricing tier, and call it simplification.
A sovereign architecture preserves the ability to respond.
Yesterday’s infrastructure decision should not become tomorrow’s permanent technical constraint.
Sovereignty Requires an Exit Door
There is a simple way to test an AI system’s sovereignty:
What happens if you need to leave?
Can you export the organization’s data and memory?
Can you preserve the adapted model?
Can you change the inference provider?
Can you move the workload from cloud to private infrastructure?
Can you replace one model without rewriting the application?
Can you continue operating if a vendor changes its terms?
If the answer is no, the system may still be private, secure, compliant, or locally hosted. Those are valuable properties.
But it is not fully sovereign.
It is dependent, just with better geography.
No Single Layer Makes You Sovereign
Not the model.
Not the infrastructure.
Not the deployment.
Not the data.
Sovereignty exists when an organization retains meaningful control across the system, along with the freedom to change it.
That is where I believe enterprise AI is heading.
Not only because regulation may demand it or security teams will insist on it, but because AI is becoming core infrastructure. Organizations will want to own the capabilities that differentiate them without permanently tying those capabilities to today’s model, vendor, or hardware.
The future of enterprise AI is not simply open.
It is not simply local.
It is not simply private.
It is portable, adaptable, governable, and replaceable.
If your AI cannot leave, it is not sovereign. It is just being hosted somewhere you recognize.
This is a conversational remix of an article I published on Backboard’s blog. Read the original deep dive here.
Top comments (1)
What’s the hardest dependency in your AI stack to replace today:
the model, memory, infrastructure, or something else?