You are only as sovereign as your weakest layer.
A government can build a domestic data center, fund a national model, and pass a sweeping AI law in a single quarter — and still not be sovereign in any way that survives a real supply shock.
Sovereignty in AI was never a trophy you win. It is a stack of dependencies — compute, data, models, applications, governance — and a set of deliberate decisions about what to own, what to govern, and what to rent.
Revised & expanded 2026 edition · Foreword by Robbie Jerrom, Red Hat
"Sovereign" has become a marketing word.
A nation can host compute domestically and depend entirely on a foreign chip supply. An enterprise can deploy "private AI" and still route its most sensitive workflow through infrastructure it does not meaningfully control. A public agency can satisfy a residency requirement while remaining captive to a vendor's roadmap, license terms, and legal jurisdiction.
The failure mode is almost never choosing the wrong dependency. It is letting dependencies accumulate without anyone having decided to take them on.
A dependency you chose with your eyes open is a strategy. A dependency you discovered during an outage is a failure mode.
Where does your data go? Who can change your models? What happens at renewal, when the provider holds the leverage? Those are not theoretical questions. They are the architecture review your procurement process probably skipped.
Five layers. One question per layer:
own it, govern it, or rent it?
Real enterprise AI is modular — assembled from bricks you own, bricks you rent, and bricks that snapped into place without anyone deciding. Nobody owns the whole box. Sovereignty does not look the same at every layer: some are achievable, others structurally concentrated. The discipline is choosing each brick on purpose.
Governance
Policy, oversight, legal frameworks, auditability, and accountability. The layer almost anyone can own — and the one that keeps the rest in line.
Applications
The services and workflows where AI actually touches the work — including the "just experimenting" ones that quietly became operational dependencies.
Models
Foundation models, fine-tuned systems, open-weight deployments, proprietary APIs. Open weights turned model sovereignty from a billion-dollar problem into a fine-tuning problem.
Data
The datasets that train, fine-tune, and feed the systems. The most achievable sovereignty layer — but residency is a fact about geography, and sovereignty is a fact about control.
Compute & Infrastructure
Chips, accelerators, networking, power, data centers. The real chokepoint — the layer almost nobody fully owns, and the one that constrains everything above it.
Owning the AI Sovereignty Stack
What nations and enterprises actually need to own in the age of AI. Revised & expanded 2026 edition.
- Why compute and the semiconductor supply chain — not models — are the real sovereignty chokepoint
- How open-weight models turned model sovereignty from a billion-dollar problem into a fine-tuning problem
- The Renewal Test that exposes which of your "strategic partnerships" are really just untested dependencies
- How the major powers actually behave — the US, China, Europe, the Gulf states, and India — and what mid-sized nations and organizations should learn from each
- The 20-minute Sovereignty Stack Self-Assessment you can run on a nation, an agency, or your own company
If you are building enterprise AI strategy, operating infrastructure, advising organisations on where to invest and where to draw boundaries: this book is the clearest thinking I have seen on the problem you are actually solving. Not the problem of capability. The problem of control.Robbie Jerrom — Senior Principal Technologist AI, Red Hat · from the foreword
How sovereign are you, actually?
The Sovereignty Stack Self-Assessment from Appendix A of the book — in a form you can put in front of a team. Rate 25 statements across five layers and see the shape of your dependencies, not just a score.
No signup required. Runs entirely in your browser — your answers never leave your machine. Which is rather the point.
The point of the exercise is not the score. It is the conversation the score forces:
Where would we be genuinely stuck if a single provider changed its terms next quarter? Which dependency did nobody actually decide to take on? Which layer would embarrass us first under real stress?
A strong total with one weak layer is not strength. It is a concealed exposure.
A book can give you the frame.
It cannot run the renewal test for you.
If your organization is working through what to own, what to govern, and what to rent — and wants help making those decisions deliberately — that is the work I do.
Executive Briefing
A focused session on the Sovereignty Stack: where the real chokepoints are, how the major powers actually behave, and what it means for your roadmap, your procurement, and your renewals. Keynote or closed-door.
Book a briefing →Sovereignty Stack Workshop
A facilitated working session built on the self-assessment: map your dependencies layer by layer, run the renewal test on the ones that matter, and leave with explicit own / govern / rent decisions and named owners.
Run the workshop →Ongoing Advisory
Standing counsel as the decisions arrive: model strategy, private AI architecture, procurement terms, exit paths, and the operational discipline that keeps "sovereign" from becoming a label you can't defend.
Start the conversation →Frank La Vigne
Frank La Vigne is a technologist, author, and longtime practitioner at the intersection of enterprise AI, data, and the organizations trying to make sense of both. He has spent his career in the rooms where AI strategy meets reality — architecture reviews, procurement negotiations, and the postmortems that follow when a dependency nobody decided on gets tested.
Owning the AI Sovereignty Stack grew out of that work: a refusal to let "sovereign" remain a marketing word, and a framework for the decisions that actually determine who controls what. He writes and speaks about AI sovereignty, data governance, and the unglamorous operational discipline that makes velocity survivable.
He does not write like someone selling a framework. He writes like someone who has sat in enough architecture reviews to know exactly where the fantasy breaks down.
The book argues from structure. The updates carry the parts that change.
Occasional dispatches on AI sovereignty as the field moves: framework updates, the examples that move too fast for print, and the renewal-test questions worth asking this quarter. No noise.
The dependencies are yours either way. The point is to make sure somebody actually does the thinking.
Tell me where you are in the stack — what you're running, what's up for renewal, what nobody decided on purpose — and I'll tell you honestly whether and how I can help.