SOVEREIGNTY · JURISDICTION

Sovereign AI Deployment: Weights, Compute, and Control in One Jurisdiction

A board that has just read its cloud contract properly wants to know what happens if the vendor, or its government, changes the terms.

Written by Jakub Rusinowski · Last updated 2026-08-01 · Hardware figures computed by our VRAM engine

This page is practitioner guidance on deployment mechanics, not legal advice — involve your counsel or DPO for decisions about your specific obligations.

Sovereign AI means the four layers of an AI system — the model weights, the compute they run on, the data they touch, and the operational control over all three — sit under a jurisdiction and an operator you have chosen deliberately. A hosted frontier API can give you contractual assurances about data residency, but it cannot give you the weights or independence from the provider's own legal environment. Open-weight models on infrastructure you own or rent under local law can give you all four, which is why sovereignty has moved from a policy talking point to a procurement criterion. For the infrastructure side of that shift, see our report on [the EU's AI Factory network](/reports/europe-ai-compute-gap-sovereignty).

Why this is hard

What the deployment gives you

You hold the weights

An open-weight model on your storage cannot be deprecated out from under you. The version you validated is the version that keeps running next year, which matters most in exactly the regulated settings where re-validating a model is expensive.

One jurisdiction, stated plainly

Hardware in your facility or an operator incorporated under local law, with no foreign parent able to compel disclosure. This is the distinction between data residency, which is about location, and sovereignty, which is about who can lawfully reach the operator.

Operational independence

Serving, updating, and monitoring performed by your own team or a local partner. No dependency on a foreign support rota, an export-control decision, or a sanctions regime you do not control.

An exit that is not a rewrite

Open weights behind a standard inference API mean swapping models is a configuration change. Sovereignty and portability turn out to be the same architectural property viewed from two directions.

The four layers, and who controls each

Sovereignty is not binary — it is a stack, and most organisations end up sovereign at some layers and not others. Mapping the four honestly is more useful than a yes/no answer, because it shows exactly which dependency to address first.

The four layers, and who controls eachYour jurisdiction, your operator. 1. Weights: Open-weight model held as a file you archive, version, and can rebuild from. 2. Compute: Owned racks or a locally incorporated operator — not a foreign-controlled tenancy. 3. Data: Corpora, indexes, and logs resident under local law and local access rules. 4. Operations: Your staff or a local partner run updates, keys, and monitoring end to end.YOUR JURISDICTION, YOUR OPERATORWeightsOpen-weight model held as afile you archive, version,and can rebuild from.ComputeOwned racks or a locallyincorporated operator — nota foreign-controlledtenancy.DataCorpora, indexes, and logsresident under local lawand local access rules.OperationsYour staff or a localpartner run updates, keys,and monitoring end to end.

What you have to satisfy

Models that fit this deployment

模型VRAM (Q4)可运行于上下文许可
Mistral Small 3.1 24B
European-published, Apache 2.0 — Where model provenance is itself part of the sovereignty case, not just where the GPUs sit.
ollama pull mistral-small3.1		15 GB24 GB GPU (RTX 3090/4090)
Mac: 24 GB 统一内存		125KApache 2.0
Qwen 3 32B
Department tier — one 24 GB card — Permissive licence and strong multilingual coverage; the pragmatic default for internal assistants.
ollama pull qwen3:32b		20.6 GB24 GB GPU (RTX 3090/4090)
Mac: 32 GB 统一内存		125KApache 2.0
GPT-OSS 120B
National / large-org tier — Frontier-adjacent quality from open weights you can archive, on two server cards.
ollama pull gpt-oss:120b		71.3 GB2×48 GB GPUs / big unified memory
Mac: 96 GB 统一内存		128KApache-2.0

Building AI capability you will still control in five years?

We help organisations and public bodies design sovereign AI deployments — mapping the four layers against your actual dependencies, selecting models with durable licences, sizing the hardware, and planning the exit from whatever you are on now. Vendor-neutral by construction: we sell no infrastructure and no models.

Ask about a workshop →

常见问题

What is sovereign AI in practical terms?
It is control over four layers: the model weights, the compute, the data, and the operations. A deployment is sovereign to the degree you can answer "who could take this away, and under whose law?" without naming a foreign entity. In practice that usually means open-weight models running on owned hardware or on infrastructure operated by a company incorporated in your jurisdiction, with your own staff holding the keys and running the updates.
Is data residency the same as sovereignty?
No, and conflating them is the most common mistake in this area. Residency is a statement about geography — your data sits in a datacentre in a particular country. Sovereignty is a statement about legal reach — which government can compel the operator to act, and which corporate parent sits above them. A foreign-owned provider running a local region satisfies residency while leaving the sovereignty question open, which is precisely why vendor country of origin became a selection criterion: 77% of organizations now factor it into AI vendor decisions, per Deloitte's State of AI in the Enterprise survey of 3,235 business and IT leaders across 24 countries.
Are open-weight models good enough to build sovereign capability on?
For the workloads organisations actually deploy — internal assistants, document analysis, retrieval over institutional knowledge, drafting, translation, coding support — yes, comfortably. A 32B-class model on a single card handles most of it, and open models in the 120B class reach frontier-adjacent benchmark territory on two server GPUs. Hosted frontier models retain an edge on the hardest reasoning tasks. The sensible posture is to pilot with your own documents and measure, rather than assuming either that open models are inadequate or that they are equivalent.
Can we get sovereignty without buying hardware?
Partly. Renting GPUs from an operator incorporated in your jurisdiction, running open weights you supply, gets you weight sovereignty, data locality, and a good deal of operational control without a capital purchase. What you do not get is independence from that operator's continued existence and terms. For many organisations that is an acceptable trade, especially as a first step — the important thing is that because you hold the weights, moving to owned hardware later is a migration rather than a rebuild.
What is the single most important thing to do first?
Archive the weights and tokenizer of any model you depend on, with a checksum, somewhere you control. It costs almost nothing, takes an afternoon, and it is the one step that converts "we use an open model" into "we can still run this if the internet, the publisher, or the licence changes." Everything else in a sovereignty programme is easier once that is true.

Go deeper