Your Weights Are a Power Station

Nobody calls a power station a convenience

When a country builds a power station, it does not describe the decision as a procurement convenience. It treats generation capacity as a strategic asset, defends it, regulates it, and refuses to let a foreign operator hold the only key. Yet when the same country wires its hospitals, courts, tax authority, and defence analysis into a large language model hosted on someone else's hardware, it files that under software licensing and moves on. That is a category error, and it is going to cost us.

Model weights are the trained reasoning capacity of an artificial intelligence (AI) system. They are not the application, not the chat window, not the friendly interface. They are the compressed result of enormous compute and curated data, the actual thing that decides, summarises, classifies, and recommends. Once a government runs material decisions through a model, those weights are doing work that used to be done by analysts, clerks, and clinicians. Strip away the branding and you are looking at infrastructure. It deserves the same seriousness as the grid, the water supply, and the ports.

Renting intelligence is renting a dependency

A security realist starts from a simple question. If this thing is switched off, withheld, or quietly changed, what breaks, and who is now in a position to break it? Apply that to a rented frontier model and the answer is uncomfortable. The weights live in another jurisdiction. The provider can revise the model overnight, deprecate it, throttle it, or be compelled by its own government to change what it does. You did not agree to any of that, and most of the time you will not even be told it happened.

We learned this lesson the hard way with energy and with semiconductors. A dependency you cannot see is the most dangerous kind, because the bill arrives at the worst possible moment. Intelligence is now in that bracket. When a national fraud-detection pipeline, a benefits-eligibility system, or a clinical triage tool depends on a model you neither host nor control, you have handed a foreign provider a lever over the daily operation of the state. That is not a partnership. It is a tenancy, and the landlord can change the terms.

People object that the alternative is slower and more expensive. Often it is. But we do not justify foreign control of the electricity grid on the grounds that imported power would be cheaper this quarter. Strategic capability is allowed to cost more than the convenient option, because the convenient option carries a tail risk the spreadsheet never prices. Renting intelligence optimises for the easy years and quietly mortgages the hard one.

Why weights are infrastructure, not just data

Three properties move weights out of the ordinary-software box. First, they are non-substitutable at speed. You cannot retrain a frontier model over a weekend if access is cut. The compute, the data pipeline, and the engineering are a multi-year build, which is exactly the profile of infrastructure rather than a tool you swap out. Second, they shape outcomes invisibly. A subtly altered model produces subtly altered decisions across millions of cases, and unlike a power cut there is no obvious moment of failure. Third, they sit upstream of almost everything else a digital state does, so a single dependency propagates into health, justice, finance, and defence at once.

That combination, hard to replace, invisible when compromised, and upstream of everything, is the textbook definition of critical national infrastructure. We already apply that label to systems with far less reach into daily decision-making than a model that quietly underwrites a million administrative judgements a week. The honest move is to apply it here too, before an incident forces the conversation. The pattern in every infrastructure failure is the same. The dependency was always there, it was simply not named until it broke.

Sovereignty is a spectrum, and most of it is fake

Here is where I get lightly contrarian. A great deal of what gets sold as sovereign AI is theatre. A foreign model running in a local data centre is not sovereign just because the rack sits on home soil. If you cannot inspect the weights, cannot run them without the vendor's permission, and cannot verify what the system actually did after the fact, you have rented a strategic capability and repainted the cage. Geography is the easy part of sovereignty. Control and proof are the hard parts, and they are the parts that matter.

Real sovereignty over intelligence means three things. You can run the model without asking anyone. You can change it without anyone's leave. And you can demonstrate, to a court, an auditor, or a citizen, exactly what it did and that nobody tampered with the record. The first two are about possession. The third is about proof, and it is the one almost everyone forgets, because possession feels like the whole answer right up until someone asks you to account for a decision and you have nothing to show them.

The regulators are arriving at the same place

This is not only a national-security argument, it is becoming a legal one. The European Union (EU) Artificial Intelligence Act brings serious obligations for high-risk systems into force during 2026, and the direction of travel is unmistakable. Operators of consequential AI will be expected to show records, demonstrate human oversight, and account for decisions rather than assert that the system was well behaved. At the same time, the long migration to post-quantum cryptography is underway, because the signatures and encrypted records we create today need to survive a future adversary with a quantum computer. Rising AI liability sits on top of all of it.

I will not invent a precise figure, but the qualitative trend is clear. The burden of proof is shifting onto whoever deploys the model, and it is shifting from policy promises toward evidence you can put in front of a regulator. If you are running rented intelligence you do not control, meeting that burden honestly is close to impossible. You cannot produce a tamper-evident account of a black box you do not own. So the compliant answer and the sovereign answer turn out to be the same answer, which is a rare and useful thing, because it means doing the right thing for security also satisfies the lawyer.

Possession is necessary, proof is what finishes the job

This is the thesis I have built Mickai around, and I will state it plainly so it does not read as an advert. Owning the weights is necessary but not sufficient. A sovereign system has to do something stronger than promise good behaviour. It has to make its behaviour checkable by someone who does not trust the vendor at all, including me. If the only assurance I can offer you is my own word, I have not built sovereignty, I have built a brand.

Mickai is a Sovereign Intelligence Operating System (SIOS), built and in production. Fifty brains, twenty-five domain and twenty-five operational, run on the Poseidon silicon substrate, and we are actively training our own models now, fine-tuning and specialising open foundations such as Llama 3.2 and Qwen 2.5 while building a sealed corpus, with funding scaling that work toward fully native weights. The portfolio behind it is real, one hundred and one filed United Kingdom patent applications and roughly two thousand two hundred and thirty-four claims, owned by Mickai LTD, with myself as the named inventor. But possession is the floor, not the ceiling.

The part that earns the word sovereign is the Open Audit Record (OAR). Every action the intelligence takes is signed before it executes, hash-chained, and append-only. The signatures use post-quantum cryptography, the United States National Institute of Standards and Technology (NIST) standard FIPS 204 (ML-DSA-65), so the record stays trustworthy as cryptography moves on. And the whole chain verifies offline, in an ordinary browser, with no trust placed in the vendor. The Pantheon chain, the one piece still in build, anchors that audit root to Bitcoin so the record sits beyond any single party's reach. You do not take my word that the system behaved. You check the record yourself, and if I tampered with it, the mathematics gives me away.

Build the power station

A nation that rents its intelligence has outsourced a strategic capability it can neither inspect nor switch off, and has done so quietly, under a procurement line that hides the stakes. The fix is not slogans about sovereignty. It is the unglamorous infrastructure work of owning the weights, running them on your own substrate, and proving what they did with a record that survives both your adversaries and your own institutions.

Treat model weights the way you treat a power station. Build it, defend it, and keep the only key. Then make the intelligence sign its name to everything it does, so trust becomes something you can verify rather than something you are asked to extend. That is the whole argument. The country that takes it seriously first will not have to relearn it during a crisis, which is the only time the lesson is ever genuinely expensive.