Sovereign AI: what do states actually control?

A closed-door breakfast during the high-level week of the eighty-first session of the United Nations General Assembly

1. Background

Sovereign artificial intelligence has become the organising ambition of national technology policy. Governments announce national compute programmes, national models, national clouds, data residency requirements and public procurement rules intended to keep strategic capability within national reach. The vocabulary now travels well beyond the states that possess the underlying industrial base, and it is increasingly used in ministerial statements by countries whose domestic capacity at most layers is close to nil.

The difficulty is that the word sovereignty is being asked to cover very different things. Artificial intelligence capability is not a single object that a State either has or lacks. It is a stack: semiconductors and specialised hardware, data centres and the energy that powers them, cloud and compute, data, foundation models, the deployment layer at which systems reach citizens, and the standards and governance layer that determines how any of it may be used. Ownership, jurisdiction and leverage differ at every one of those layers. A State may regulate deployment while owning no compute. It may host data centres it does not control. It may procure models it cannot inspect. It may hold a seat in a standards body without the technical staff to use it.

The work of Professor Corneliu Bjola on the digital stack offers a way to move this conversation from slogan to inventory. Rather than asking whether a State is sovereign in artificial intelligence, it asks a more useful question at each layer: what is owned, what is hosted, what is regulated, what is procured, and which instrument of foreign policy actually reaches it.

That question cannot be answered by a strategy document. It requires officials able to conduct the inventory and to negotiate on its basis. This is the commitment at the centre of the Samarkand Declaration on Tech Diplomacy, adopted on 1 November 2025, under which signatory States appoint Tech Ambassadors or equivalent senior tech diplomats, including digital, cyber and AI ambassadors and special envoys for technology, and build national capacity in tech diplomacy. Implementation will be reviewed at the Tech Diplomacy Asia Symposium in Tashkent on 14 October 2026 and at the Tech Diplomacy Global Assembly in Paris on 11 December 2026.

2. The breakfast

The Tech Diplomacy Global Institute and Mia AI convene a closed-door breakfast in New York on the morning of Thursday 24 September, during the high-level week of the eighty-first session of the General Assembly. It follows the roundtable on the implementation gap held the previous afternoon and takes the question one layer down, from who delivers governance to what governance can reach.

The format is deliberately small. Around twenty participants, seated at one table, under the Chatham House Rule, with no press and no podium. High-level week is saturated with panels; what it lacks is unrecorded conversation among people who can act. The session concludes before the day’s proceedings at United Nations Headquarters begin.

3. Objectives

• To test sovereign artificial intelligence claims against the layers of the digital stack, distinguishing what States own, what they host, what they regulate and what they procure.
• To identify the layers at which diplomatic instruments actually bite, including export controls, public procurement, standards bodies, bilateral technology partnerships and cloud sovereignty arrangements.
• To hear from States, in a setting where officials can speak without attribution, which dependencies they consider acceptable and which they consider strategic.
• To connect the diplomatic layer with the institutional one, linking the appointment of Tech Ambassadors or equivalent senior tech diplomats to the technical capacity required to conduct a national inventory.