BLUF — Rented AI can cost you twice — once in fees, once in the proprietary edge that leaks out with every prompt. Own your data, your model choice, and the decisions your agents make: run domain-tuned, open-weight models you control, and keep the intelligence your operation generates. Sovereignty is an architecture choice, not a build-it-yourself project.
Three things you can own — or accidentally rent
“Sovereign AI” sounds like a policy term. For an operations leader it’s concrete, and it comes down to three layers you either control or quietly hand away.
Your data. Where does your operational data physically go, and who can see it along the way? For PHI, clinical-trial records, or supplier contracts, this isn’t a preference — it’s law and contract. If sensitive data leaves your environment to reach a model, you’ve created exposure you may one day have to explain to someone who isn’t sympathetic.
Your model. If your entire operation runs on one vendor’s model, you’ve inherited their pricing, their outages, their deprecations, and their policy changes. The model should be a component you can choose and swap — not a foundation you can’t move off without rebuilding everything on top of it.
Your accumulated intelligence. Every decision your agents make, every escalation, every judgment you encode is institutional knowledge. If it accrues inside a system you don’t control, you’re building someone else’s asset with your operation’s hard-won experience.
This isn’t abstract. In a governed procurement operation we run, the same team now handles the volume three times — and the durable advantage there isn’t the model, which any competitor can license next week. It’s the accumulated record of how that operation learned to decide: which suppliers slip, which clarifications threaten the schedule, which exceptions are real. Swap the underlying model tomorrow and that intelligence stays — because the architecture kept it ours.
The model is rented. The moat is owned.
Why this is an operations issue, not an IT footnote
Operations leaders already manage single points of failure everywhere — one supplier, one plant, one key person. A single-model, single-vendor AI dependency is the same risk in a new place, and it usually sits outside the risk register because it arrived as “an AI tool,” not “a core dependency.”
The regulated angle makes it sharper. A health system moving patient data across a border, a life-sciences firm under GxP and data-residency rules, a supply-chain operation whose contracts prohibit sharing pricing with third parties — for all of them, “where does the data go and who trains on it?” is a board-level question. The concentration risk and the compliance risk are the same conversation.
Sovereign doesn’t mean building it all yourself
Here’s the misconception that stops leaders cold: they hear “sovereign AI” and picture training a foundation model in-house, an impossible project for an operations team. That’s not it.
Sovereignty is an architecture choice, not a build-everything mandate. It means your data stays in your environment, sensitive fields are redacted before anything leaves, the model is a swappable component you can change when a better or cheaper one appears, and the record of every decision lives with you. You can use the best models on the market and still be sovereign. The difference is whether you’re a tenant of the intelligence layer or its owner.
Think back to the cab analogy. Uber owns the intelligence layer; the drivers are interchangeable supply feeding the platform. In your own operation, make sure you’re the platform — not the interchangeable driver handing your proprietary operational data to someone else’s system to make it smarter.
Right model, right job — not the biggest one for everything
There’s a quiet assumption in a lot of AI plans: bigger model, better results. In operations that’s usually wrong — and expensive. The largest, most capable models are impressive, but they’re also the slowest and priciest to run, and they’re the ones you’re least able to keep inside your own walls. The skill isn’t picking one model. It’s matching the model to the work.
Think of your workload in layers. The bulk of it is high-volume routine work — reading a document, sorting a request, pulling three fields, passing it on. That work doesn’t need a genius; it needs something fast, cheap, and reliable. A small, efficient model handles it well — and it’s small enough to run in your own environment. Reserve the large, expensive frontier model for the rare, genuinely hard judgement calls where its extra power actually earns its cost.
This is a sovereignty decision, not just a budget one. The small, efficient models are exactly the ones you can run yourself — so the more of your volume you serve with them, the more of your operation stays under your control, and the less of your data has to travel to someone else’s model to get the job done.
And size isn’t the only reason to reach for a model you own. A model tuned on your domain — your claims history, your supplier base, your regulatory rules — will often out-perform a bigger, general-purpose model on your actual work, because it knows your world instead of the whole internet. The best place to build that domain intelligence is an open-weight model you control: you teach it on your data, that data stays in-house, and the sharper it gets, the more the advantage is yours to keep rather than rented back to you. This is exactly the “weights and alpha” Karp warned about not giving away — except here you’re keeping them. For a healthcare or defence-procurement operation, the model that’s cheapest to run, safest to own, and best at your work is frequently a compact, domain-tuned, open one.
Open weight or closed weight — a rented brain or an owned one
The second choice decides whether owning your intelligence is even possible. Models come in two kinds, and the difference is simply where the “brain” lives.
A closed-weight model is one you can only reach through the vendor’s service. The model itself — the “weights,” the part that makes it work — stays locked in their data centre. You send your data out to it, you can’t look inside it, and you can’t run it on your own. It’s a rented brain: convenient, often very capable, but living in someone else’s building on someone else’s terms.
An open-weight model is one whose brain is published, so you can download it and run it on your own computers. Your data never has to leave, you can inspect and tune it, and you can even run it fully disconnected — air-gapped — for your most sensitive work. It’s an owned brain.
Sovereign doesn’t mean going all-open or all-closed. It means you get to choose — running the sensitive, high-volume work on models you control, and reaching for a closed frontier model only when a specific job truly needs it. The danger was never using a closed model. It’s having no other option.
What to do this week
You don’t need a sovereignty strategy to start. You need to know where you actually stand.
• Trace your data’s passport. For one AI-touched workflow, map exactly where the data travels: what leaves your environment, to whom, and in which jurisdiction. If you can’t draw the map, that gap is the finding.
• Ask the exit question. Of any AI vendor: “If I need to switch models or providers next year, what happens to my data, my configuration, and the intelligence we’ve built?” A sovereign architecture has a clean answer. Lock-in has an awkward silence.
• Right-size before you scale. Take your highest-volume workflow and ask whether a smaller model you could run yourself would do the job — instead of sending every item out to a frontier model. Matching the model to the task is where cost control and sovereignty meet.
• Separate the model from the moat. Decide what you’re comfortable renting — raw model capability — and what you must own: your data, your decision records, your workflow logic. Write that line down before the next pilot, not after.
A final thought
The leaders getting this right treat AI capability like electricity — useful, buyable, and swappable — and treat their operational intelligence like the business itself: not something you’d ever hand to a landlord.
And this only gets bigger as you scale. The more of your operation runs on governed intelligence, the more that intelligence is the operation — so scaling on a layer you don’t own means compounding someone else’s advantage instead of your own.
When Palantir’s Karp and Microsoft’s Nadella — one building for governments, the other running the world’s biggest enterprise cloud — land on the same warning from opposite ends of the industry, the move isn’t to panic. It’s to ask their question of your own operation, before the next pilot rather than after.
So here’s the one to sit with: everything your operation is learning right now — the decisions, the exceptions, the hard-won judgment — whose asset is it quietly becoming, yours or your vendor’s? If you can’t answer that with a confident “ours,” sovereignty is the conversation to have first. Tell me in the comments where your operation stands.
Recent blogs
Secure your agents
We’d love to chat with you about how your team can secure and govern Ai agents everywhere
