The EU buys sovereign cloud, and the AI layer needs the same logic

The logic behind the European Commission's sovereign cloud procurement applies with even greater force to AI: if resilience and control require avoiding dependence on any single infrastructure supplier, then the models that reason over an institution's most sensitive data need the same independence. On 17 April 2026 the Commission awarded 180 million euros in sovereign cloud contracts to four providers for EU institutions, deliberately diversifying to avoid reliance on one supplier and to strengthen the Union's digital sovereignty posture. We think that decision is right, and we think it is incomplete on its own, because sovereignty that stops at the infrastructure layer leaves the intelligence layer exposed.

We build Mickai, a Sovereign Intelligence Operating System that runs entirely on an organisation's own hardware, on premise and air-gapped where the mission requires it. Watching the EU buy sovereign cloud, we see a procurement principle that public bodies everywhere can extend upward: own the layer that thinks, not just the layer that stores.

What did the European Commission decide on 17 April 2026?

The Commission awarded 180 million euros in cloud contracts to four providers to serve EU institutions, and it structured the award deliberately so that no single supplier could become a dependency. According to the Commission's own announcement, the procurement forms part of a wider effort to strengthen the Union's digital sovereignty, treating resilience and control as requirements to be engineered into contracts rather than hoped for afterwards. The signal to every public body in Europe is straightforward: concentration is a risk, diversification is a policy instrument, and sovereignty is now something you specify in a tender.

Why does sovereignty matter even more at the AI layer than at the cloud layer?

Because the AI layer sees the meaning of the data, not just the bytes. A cloud provider hosts files and workloads; a model reads case files, draft legislation, procurement plans and citizen records, and produces judgements about them. If that model runs as a remote service, every prompt and every response transits infrastructure the institution does not control, and the dependency the Commission just spent 180 million euros engineering out of the storage layer reappears one layer up, in a more concentrated form. A single external AI provider becomes a single point of exposure for the most sensitive reasoning an institution does, and a single point of failure for every workflow built on top of it. Diversifying cloud suppliers while consolidating all machine reasoning onto one external service would be sovereignty in the basement and dependency in the boardroom.

What does applying the same logic to AI look like in practice?

Applying the Commission's logic to AI means three things: run the models on infrastructure the organisation owns, remove the single supplier from the reasoning path, and keep the evidence of every action under the institution's own control. That is the design brief we set ourselves when we built our operating system. In practical terms, a sovereign AI layer should have the following properties.

• It runs on the organisation's own hardware, on premise, with fully offline operation available so that nothing needs to transit a third party.
• It can operate air-gapped, so the environment that reasons over sensitive data has no route out of the building.
• No single model sits alone in the reasoning path. Our cooperative multi-model consensus substrate requires specialist sovereign models to agree before any sensitive action runs.
• Control is anchored in a hardware-held root of trust that the institution physically possesses, not in a remote provider's account.
• Every action is cryptographically signed to the Open Audit Record, which is post-quantum secure, tamper-evident and verifiable offline.
• Sensitive actions are gated by voice-biometric confirmation, binding authorisation to a person rather than only to a credential.

“Sovereign cloud answers the question of who holds your data. A sovereign intelligence operating system answers the harder question of who controls what reasons over it.”

How does diversification work inside one operating system?

Diversification inside our operating system works the way the Commission's procurement works across suppliers: no single participant is trusted alone. Our cooperative multi-model consensus substrate routes sensitive work across specialist sovereign models that must agree before an action proceeds, so a weakness or failure in one model does not silently become the institution's decision. The Commission spread its award across four cloud providers so that no one supplier could become a chokepoint; we apply the same principle to inference itself. And because these are our own models running on the customer's hardware, diversification here does not add external dependencies, it removes the last one.

What should procurement teams ask of an AI supplier now?

The first question is whether the supplier can operate entirely inside the institution's perimeter, because every other assurance follows from that. A procurement team taking its cue from the Commission's example should ask where the models run, who can reach the data while they run, what evidence exists of each action afterwards, and what happens if the supplier disappears. Our answers are that the operating system runs on the customer's hardware, that data never has to leave the environment, that the Open Audit Record provides a signed, offline-verifiable account of every action, and that a hardware-held root of trust keeps ultimate control with the institution rather than with us.

Scope matters to procurement as well, because sovereignty that covers one narrow use case forces the institution back into dependency for everything else. We launch with ten production-ready studios out of 87 on one operating system, with the remaining 77 in development, so an organisation can begin with the workloads that matter most and expand without ever changing its sovereignty posture. The engineering underneath is the subject of 104 filed UK patent applications across 2,340 claims, though we would rather be judged on the architecture than on the paperwork.

The EU has shown that sovereignty can be specified, procured and enforced at the infrastructure layer. The next tender should specify it at the intelligence layer too, so that institutions own a capability rather than rent a dependency.

Frequently asked questions

What did the EU's sovereign cloud award actually cover?

On 17 April 2026 the European Commission awarded 180 million euros in sovereign cloud contracts to four providers to serve EU institutions, deliberately diversifying across suppliers to avoid dependence on any single one and to strengthen the Union's digital sovereignty posture.

Is sovereign cloud enough to make AI sovereign?

No. Sovereign cloud governs where data is stored and processed, but an AI model consumed as a remote service still carries prompts, responses and reasoning outside the institution's control. Sovereignty at the AI layer requires the models themselves to run on infrastructure the organisation owns, offline and air-gapped where needed.

Can AI genuinely run fully offline?

Yes. Our operating system runs entirely on the customer's own hardware and operates fully offline, air-gapped where required. Our own sovereign models run locally, and the Open Audit Record is verifiable offline, so neither operation nor audit depends on any connection to an external service.

How does a multi-model consensus substrate reduce supplier risk?

It removes the single model, and therefore the single supplier, as a point of failure. By requiring specialist sovereign models to agree before any sensitive action runs, the substrate applies the same diversification principle the Commission used across cloud providers to the reasoning layer itself.

What is MICKAI?

MICKAI is a Sovereign Intelligence Operating System (a SIOS) that runs on the customer's own hardware, on premise and air-gapped, with every action cryptographically signed to the post-quantum secure, tamper-evident Open Audit Record. It comprises 87 studios on one operating system, with ten production-ready at launch and 77 in development, and its engineering is covered by 104 filed UK patent applications across 2,340 claims, filed rather than granted.