The Case for Sovereign AI: Own It, Don't Rent It

Sovereign AI is artificial intelligence that an organisation owns and runs on infrastructure it controls, so that its data, its models and its records never leave its custody. You should own it rather than rent it whenever the work is regulated, because renting AI from the cloud means renting it on someone else's terms: your sensitive data lives on their machines, under their jurisdiction, readable under their process, and available only while they choose to serve you. Ownership reverses all of that. For a bank, a hospital, a law firm or a public body, ownership is not the paranoid option. It is the honest default.

• Sovereign AI runs on infrastructure you own and control, keeping data, models and audit inside your custody.
• Renting from the cloud trades custody, provability and predictable cost for elasticity and convenience.
• For regulated work, custody and provable evidence usually matter more than burst scale.
• The natural buyers are banking, healthcare, legal and the public sector.
• Ownership is the honest default when you must be able to prove, offline, what your AI did.

What sovereign AI actually is

Sovereign AI is a system where the intelligence comes to your data instead of your data going to the intelligence. The model runs on hardware you control. It does not depend on a live connection to a vendor to function. And it keeps a record of its work that you can verify yourself. That is a stricter definition than the marketing term sovereign cloud, which often just means a regional data centre still operated by someone else.

I test any sovereignty claim against three questions. Can you own it, meaning does it run on your infrastructure with the model in your hands? Can you run it offline, with no outbound connection required? And can you prove what it did, independently, without trusting the vendor? If the answer to any of those is no, you are renting, whatever the label on the invoice says.

The trade you are really making

Renting AI from the cloud is genuinely attractive on paper. You get elasticity, so you can burst to enormous scale for a spike and back down again. You avoid large upfront investment. Someone else runs the operations, patches the servers and upgrades the models. For a great many businesses that is the right answer, and I would not pretend otherwise.

But that convenience is paid for in three currencies that regulated organisations cannot easily spend: custody, provability and predictable cost. You give up custody of your data, you inherit a chain of trust you cannot fully inspect, and you accept a metered bill that scales with use. Sovereign AI makes the opposite trade. You take on the responsibility of running the system, and in return you keep custody, you gain evidence, and your costs stop being a function of how many questions your staff dare to ask.

Custody: whose machine holds your data

Custody is the heart of it. When your case files, patient notes or transaction records are processed on a third party's servers, they are subject to that party's jurisdiction, its staff, its legal process and its incidents. Encryption in transit and at rest helps, but the data still has to be decrypted to be used, and at that moment it is on a machine you do not control.

With sovereign AI the data never makes that journey. Mickai brings its engine to the data, running a capable model locally through Poros, bound to loopback only and fully offline. The sensitive material stays where it already lives, behind your existing controls. There is no outbound path to worry about because there is no outbound path.

Provability: evidence, not assurances

Keeping data local is necessary but not sufficient. In regulated work you must be able to show what happened: which record was accessed, what the system did with it, and that the log has not been altered after the fact. A screenshot and a vendor's word will not survive an audit or a courtroom.

This is why Mickai writes material actions to the Open Audit Record. The OAR is post-quantum signed under FIPS 204 (ML-DSA-65) and verifiable offline. It is tamper-evident, which is the honest claim: not that a record can never be interfered with, but that interference is detectable independently, by you, without a network and without trusting us. Turning compliance from assurance into evidence is one of the topics I have written up in technical preprints on Zenodo.

Cost you can predict versus cost you rent

There is a financial dimension too, and it is not just about the sticker. Rented AI is metered. The more useful it becomes, the more your people use it, and the larger and less predictable the bill grows. That creates a quiet tax on curiosity, where teams ration their own use of a tool that is supposed to help them.

An owned system behaves like an owned asset. Once it is running on your hardware, using it more does not multiply a per-query charge. For organisations that need to budget years ahead, and that expect heavy internal use, a predictable owned cost base is often easier to defend than an elastic rental that rewards you for asking fewer questions. I will not quote figures here, because the honest answer depends on your estate, but the shape of the trade is clear.

Who sovereign AI is for

Sovereign AI is not for everyone, and I think it is more credible to say so. If your data is not sensitive and your work is not regulated, the cloud is probably the pragmatic choice. Sovereign AI earns its keep where the stakes are custody and evidence.

That means banking and financial services, bound by supervisory expectations on data and outsourcing. It means healthcare, holding special category data under strict rules. It means legal work, carrying privilege that cannot be casually exported. And it means the public sector, where national and citizen data should not sit by default on infrastructure owned elsewhere. These are the organisations Mickai is built for, and the ones joining the first cohort of design partners.

Why ownership is the honest default

The industry has quietly made renting the default and treated ownership as the exotic, expensive exception. For regulated work I think that is backwards. When you are legally accountable for data you cannot afford to lose control of, the default should be the arrangement that keeps you in control, and renting should be the thing you justify.

That is the case for sovereign AI in one line: own it, do not rent it. Own the system, run it offline, and prove what it did. Mickai is my attempt to make that not just possible but practical, on one engine with sixty-three studios, fourteen production-ready at launch and forty-nine in active development. If it fits your world, the closed beta is open at mickai.co.uk/beta.

Frequently asked questions

What is the difference between sovereign AI and sovereign cloud?

Sovereign cloud usually means a regional data centre that is still operated by a third party. Sovereign AI means the system runs on infrastructure you own and control, with the model in your hands, able to work offline and to prove its actions independently. One relocates the vendor's cloud; the other removes the dependency.

Is owning AI not far more expensive than renting it?

Not necessarily, and the comparison depends on your estate and how much you use it. Rented AI is metered, so cost grows with use and is hard to predict. An owned system behaves like an asset with a predictable cost base, which regulated organisations that expect heavy internal use often find easier to plan around. I do not publish prices, because the honest figure is specific to each deployment.

Can a local system be as capable as the cloud?

For the work most organisations actually do, yes. Mickai's Poros engine runs a large, capable model on local hardware, and sixty-three studios sit on that one engine, fourteen production-ready at launch. Offline does not have to mean underpowered.

How does sovereign AI help with compliance?

By producing evidence rather than assurances. Mickai records material actions to the Open Audit Record, which is post-quantum signed under FIPS 204 (ML-DSA-65) and verifiable offline, so an auditor or regulator can confirm what happened without trusting the vendor or needing a connection. It is tamper-evident, meaning interference is detectable.

Which industries is sovereign AI for?

Primarily regulated ones: banking and financial services, healthcare, legal, and the public sector. These are the settings where custody of data and provable evidence of what the system did matter more than the elastic scale of the cloud.

Who builds Mickai and how do I try it?

Mickai is built by Mickai LTD (Companies House 17166618), and I am its founder and named inventor, Micky Irons. The closed beta is open at mickai.co.uk/beta, where we are onboarding a first cohort of design partners.