Sovereign AI
Compute, data centres and the power to run them have become national infrastructure. Every capable state is now buying the hardware. Few are buying the capability. Our work separates the programmes that build sovereignty from the announcements that perform it.
The sovereign stake
Compute agreements signed in 2026 will set dependency structures for a decade. Export terms are volatile: Washington rescinded the AI Diffusion Rule before it took effect, and the GAIN AI Act now before Congress could tighten access again. Procurement must price in revocation risk. The EU's gigafactory call, the UK's push from 21 towards 420 AI exaFLOPS, and Gulf deliveries landing this year mean the window for favourable terms is open now.
Grid connection lead times, not chip supply, will decide which national programmes actually energise. Ministers who negotiate this as infrastructure procurement, not technology theatre, will sign terms that survive the next export-control cycle. The Institute does not duplicate the convening states already run themselves; it is the independent capability to bring in once a procurement decision has to be made.
The board-level stake
Sovereign requirements are now a procurement gate, not a talking point. Regulated sectors and Gulf and EU public tenders increasingly demand in-country processing, auditable model supply chains and demonstrable exit routes from hyperscale providers. Boards should expect data-residency clauses to fragment their AI estate across sovereign regions, each with its own pricing and capability lag. Chief AI officers need three things on paper: where every model runs, what breaks if a US export decision changes, and how fast workloads can move. Enterprises that can answer those questions will win government work. Those that cannot will be quietly disqualified from it.
Sovereign AI stopped being rhetoric and became a budget line. NVIDIA booked over $30 billion of sovereign revenue in its last fiscal year, roughly triple the year before. Abu Dhabi broke ground on Stargate UAE in March 2026, with a 200-megawatt first phase due in the third quarter. HUMAIN's first Riyadh and Dammam facilities come online this year. Europe, by contrast, has €20 billion earmarked for four to five gigafactories and has yet to close a formal call for sites. The pattern is uncomfortable but clear: the Gulf executes, Europe deliberates, and nearly everyone buys from the same two or three vendors. Owning racks of GPUs is not the same as owning capability.
Sovereignty is a stack, not a purchase: energy, chips, compute, models, data, talent, and the institutional capacity to operate and audit what runs on top. No state owns all of it; even the United States imports its lithography. Real sovereignty is therefore negotiated dependence with reversibility built in. The test is simple. If a supplier or an ally changes the terms, what breaks, and how fast can you re-route? Washington rescinded the AI Diffusion Rule in May 2025; the GAIN AI Act would partially resurrect it. Access terms shifted twice in eighteen months. Any national strategy that assumes stable chip access is not a strategy. Contracts, second sources and tested exit routes are the real sovereign assets.
Much of what is sold as sovereignty is theatre. A national foundation model that no ministry deploys is a press release with a training bill. Our founder coined the term Complex AI for the discipline this actually demands: AI operating inside complex, high-stakes, multi-objective systems (benefits, borders, grids, hospitals) where every decision must be traceable and auditable. By that standard, sovereignty is measured at the point of governing, not the point of the demo. Can the state inspect the weights? Reconstruct a decision for a court or a citizen? Switch providers without losing a decade of operational data? The scarce resource is not GPUs. It is evaluators, operators and procurement officers who can hold a vendor to account.
For small and mid-sized states, sequencing beats scale. Start with data governance and residency, because national data is the only input you uniquely hold. Build inference capacity before training capacity; most sovereign workloads are inference. Fine-tune and distil open-weight models on national data before funding a frontier run. Contract for training compute with exit clauses and portability tested, not promised. And anchor power early: with more than 2,500 gigawatts of projects stuck in grid queues worldwide (IEA, Electricity 2026), electricity, not silicon, is the binding constraint of 2026. A state that follows this order buys capability at every step. A state that starts by announcing a national model usually ends with a benchmark score and a dependency.
And the infrastructure beneath all of it is now treated as what it is: critical national infrastructure, in the same category as power and water. That reframing has consequences. Five gigawatt-scale data centres come online in 2026, each drawing the output of a large nuclear reactor, even as North America's grid regulator warns of electricity shortfalls from this summer.
Sovereign wealth funds, many of them Gulf, have become the buildout's anchor investors. And compute is no longer only an economic asset: in March 2026, drones struck hyperscale data centres in the UAE and Bahrain, the first time AI infrastructure became a kinetic target. A sovereign strategy that does not treat its data centres as infrastructure to be powered, hardened and defended is not a strategy. It is a wiring diagram waiting for a bad day.
- What must a state actually own, and what can it safely rent, before its AI can be called sovereign?
- How should governments write compute and model contracts so that access survives a policy change in Washington?
- When does a national foundation model earn its cost against fine-tuned open-weight alternatives, and who measures that honestly?
- How should small and mid-sized states sequence data, inference, energy and talent to buy capability rather than theatre?
- What traceability and audit standards make a sovereign AI system fit to govern, not just to demo?
Fellows for this frontier are being appointed.
People who have built, governed or operated real systems in this domain. The first five fellows are named, and the rest of the founding cohort follows in September.
Have you done this at scale in Sovereign AI? We want to hear from you.
Put yourself forward