Sovereign AI compute: what needs settling before a national program signs.

Written for the ministries, agencies, national research bodies and their advisors running a first sovereign AI compute procurement — and for the operators and developers being asked to supply one.

The shape of it is familiar. A ministry has budget approved, a site shortlisted and a mandate to be running by the next fiscal year. Procurement rules require an open process measured in months. The capacity it actually wants is being allocated this quarter, to whoever can commit first. Nothing about that is unusual, and nothing about it resolves itself.

National and public-sector programs are now among the largest single buyers of training capacity, and most are running a procurement they have never run before. The hardware is familiar. What sovereignty adds — jurisdiction, control, continuity and public accountability — changes almost every term around it. These are the ones worth settling first.

Jurisdiction is a specification, not a preference

Where the hardware physically sits, under whose law the operator is domiciled, and where the data rests are three separate questions with three separate answers. A cluster inside the country under a foreign-owned operator is a different asset from one under a domestic operator, and both are different from owned hardware in leased space. Decide which of these the program is actually buying before the shortlist is drawn, because it narrows the market more than any other single decision.

Own, lease, or something between

Enterprise buyers mostly lease: GPUaaS or a dedicated cluster on a term. Sovereign programs more often land on owned hardware in leased or purpose-built capacity, or on build-to-suit, because control over the asset is the point. That choice trades speed and capital for continuity. It also moves the program into a longer timeline, which makes the next item the one that decides the calendar.

Power is the schedule

For a program of any scale, the energisation date set by the operator and the utility is the binding constraint, not available space and not GPU supply. Whether a signed interconnection agreement sits behind that date or a study result is the first diligence question. A national program usually has more leverage with the utility than any private buyer and less habit of using it. Power procurement belongs inside the program, not downstream of it.

Export controls and supply

Advanced accelerators are subject to export-control regimes that vary by jurisdiction and change with little notice. The United States Bureau of Industry and Security maintains controls on advanced computing semiconductors and updates them periodically, and other jurisdictions maintain their own. The status of a specific platform for a specific destination should be confirmed with counsel before a commitment is made, and the supply timeline depends on it. This is not a formality to clear at the end. It shapes which hardware, from which channel, on which date.

Procurement process versus market speed

Public procurement cycles are longer than allocation windows. Capacity that is available in a given quarter will not wait for a tender to close. The structure that works is one where a qualified position can be held while approval runs — which raises the question of deposits in a public context, and who is able to place one, on what authority, and under what refund terms. At our desk, build-to-suit clusters with committed ready-for-service dates currently carry a deposit as the entry condition, and nothing is held without one. Settling that early is the difference between a program that lands the capacity it selected and one that lands what was left.

Counterparty and continuity

The twelve diligence questions for capacity apply in full, with one addition. Who owns the site, how the build is financed and what happens if a private operator restructures are ordinary questions for an enterprise. For a program carrying a national asset they are continuity questions, and the answer belongs in the contract as step-in rights and assignment terms rather than in reassurances.

Security written in, not alongside

Personnel clearance, on-site access, physical and network segregation, and audit rights are specifications. They should sit in the contract with the density and the fabric, not in a side letter that a later operator may not have seen. An operator who can deliver these would rather have them written down than negotiate them after the fact.

The program is not buying GPUs. It is buying the right to keep running them, in its own jurisdiction, on its own terms, for the life of the contract.

Asked about this

What makes sovereign procurement different from an enterprise GPU purchase?

Jurisdiction, control and continuity. Where the hardware sits, under whose law the operator is domiciled, whether the program owns or leases, and what happens to a national asset if a private operator restructures. Public timelines also run longer than allocation windows.

Should a program own the hardware or lease capacity?

Programs that prioritise control more often land on owned hardware in leased or purpose-built capacity, or on build-to-suit. That trades speed and capital for continuity, and moves the schedule onto power energisation rather than GPU supply.

How can a program hold capacity while a tender is running?

By settling early who can place a deposit, on what authority, and under what refund terms, so a qualified position can be held against a committed ready-for-service date while approval runs.

Sources

• US Bureau of Industry and Security — export controls on advanced computing semiconductors.
• NVIDIA GB300 NVL72 — current rack-scale platform and its power and cooling profile.
• The Chief Negotiators, build-to-suit availability — committed ready-for-service dates, term and deposit conditions seen at our desk.
• Where AI data center capacity is actually available — market-by-market power position.

Related: powered shell & build-to-suit, data center capacity sourcing, and power procurement.