“Sovereign AI” has become the default phrase in national AI plans from Abu Dhabi to Ankara. The UAE is building a 1 GW national cluster with Stargate UAE, Türkiye’s 2026–2030 plan targets 1 GW of domestic capacity, and Egypt’s second national AI strategy puts compute and data at its centre. But sovereignty is not one decision. It is a set of decisions, one per layer of the stack.
Five layers, five decisions
| Layer | Sovereignty question | Typical choice |
|---|---|---|
| Power and data centres | Can the service run if the grid or a foreign operator fails? | Partner with a national operator; own only at national scale |
| Compute (GPUs) | Who decides who gets capacity? | Mix: reserved national capacity plus commercial in-country cloud |
| Models | Can we inspect, adapt and keep running the model? | Open-weight models adapted locally; licensed models hosted in-country |
| Data and knowledge | Where do records live and who holds the keys? | Own. This is the non-negotiable layer |
| Applications and operations | Who can change, audit or stop the system? | Own the platform and the audit trail; buy or partner for build capacity |
Own the data and the off switch
If a government owns only one layer, it should be data and operations: the records, the encryption keys, the logs and the authority to change or stop a system. Everything else can be bought or partnered without giving up control, provided the contract and architecture allow you to move.
Models: adapt, don’t start from zero
Pre-training a frontier model costs more than most national AI budgets and rarely improves public services. What does improve them is a model that understands the national language, dialects and administrative terms. Open-weight models, fine-tuned and evaluated on local data and hosted in-country, deliver most of that value at a fraction of the cost. Keep the evaluation sets: they are what let you switch models later.
Compute: plan for both capacity and continuity
National clusters solve scale. They do not remove the need for planning per workload: training, batch inference and real-time services have different latency, cost and availability needs. The usual pattern is reserved capacity for critical services, commercial in-country cloud for the rest, and a tested plan for moving workloads between them.
Partners: what to ask
- Can we run the system without you after the contract ends?
- Do logs, prompts and outputs stay in-country, and can we prove it?
- Which parts are open, which are licensed, and what happens on licence change?
- Who is on call at 3 a.m., and in which jurisdiction?
How GridStudio works
We are provider-neutral. We map each layer, recommend what to own, buy or partner, and then build and run the system with in-country hosting and a full audit trail. See AI strategy, AI infrastructure and governance and compliance.
Questions
AI that a country or organisation controls: where the data and models run, who can access them, which law applies and whether the service keeps running if a foreign supplier withdraws.
No. Few governments need to pre-train a frontier model. Most sovereignty comes from controlling hosting, data, access and operations, and from adapting open-weight models to national languages and rules.
Usually the data and the operations layer: where records live, who holds the keys and who can switch the system off. Compute ownership matters for scale and continuity, but it is not the only path to control.