Sovereign AI

Sovereign AI: what the words mean, who sells it, and what a state actually buys. Every sentence below is the book’s own, from chapter 13 of THE AI COMPUTE RACE, A Change of Landlord, where the book follows the buyers who wanted the machine on their own soil: Europe, Australia, Canada, the Gulf and the G20’s ministers.

The seller’s definition

NVIDIA had put a definition of sovereign AI on the record before they met. It came in a post of 28 February 2024 over the name of Keith Strier, its vice president of worldwide AI initiatives, whom the company describes as responsible for its “strategic and commercial engagements with foreign governments.” The post says the term “refers to a nation’s capabilities to produce artificial intelligence using its own infrastructure, data, workforce and business networks.” “From nothing the previous year, we believe sovereign AI revenue can approach the high single-digit billions this year.” On the call of 25 February 2026 she said that in the fiscal year just ended the sovereign business “more than tripled year-over-year and over $30 billion, driven primarily by customers based in Canada, France, the Netherlands, Singapore and the UK.” Six months later she said through whom it was sold. On its vendor’s own accounts, sovereign AI is an American product line sold through regional partners, and a country buying it buys the landlord’s chips with a local address on the lease.

A sceptic’s question

She said in an interview published on 19 March 2026 that “to be totally honest, I don’t always know what people mean when they say sovereign AI.” She offered “the version of it that makes sense to me,” countries thinking in the aftermath of how the internet and big tech developed and how dominant American companies have been. She also said that “I just don’t think that the goal of sort of full independence makes sense.” The question she asks instead is which particular applications a country needs to run independently. Her question is the buyers’ real one, and it has a second half the vendors do not put: whose chips those applications will run on.

Europe’s version

The framework Brussels and Washington published on 21 August 2025 says that the European Union “intends to purchase at least $40 billion worth of U.S. A resolution of 22 January 2026 on European technological sovereignty and digital infrastructure records among its premises that the Union “relies on non-EU countries for over 80 % of digital products, services, infrastructure and intellectual property” and that “most data centres in Europe are not owned by European companies.” The European Commission’s own staff named who uses the capacity, in the impact assessment they published on 3 June 2026. On 3 June 2026 it adopted a proposal for a Cloud and AI Development Act, which would set “a Union cloud computing sovereignty framework comprising four Union assurance levels” for the cloud services the Union’s public bodies buy, each activity assigned its level by a risk assessment. For a state buying sovereignty by the address of its servers, that is the lesson in one case. Europe has not left and its landlord has not changed; its sovereign program is a purchase order made out to him.

The public models, and whose chips they ran on

Switzerland’s public model, Apertus, was “trained on the Alps supercomputer at CSCS in Lugano,” its makers said when they released version 1.5 on 24 July 2026, a release they called a step toward “building a sovereign AI infrastructure.” Their technical report, posted in October 2026, puts both of the release’s models on NVIDIA’s Grace Hopper modules, “a flexible reservation of GH200 nodes that grows from approximately 150 to 600 nodes as production ramps up.” A continent that holds the chokepoint tool still buys its chips from America.

The G20’s ministers

The ministers then adopted, by consensus, a statement calling on members “to develop their own policies and preserve national sovereignty in the governance of emerging technologies” and stating that they “aim to build our own sovereign policy frameworks informed by the Carolina Principles for Emerging Technologies.” The statement adopted on 2 September claimed sovereignty over the rules for the visiting delegations and their host, while the layers on offer came from American firms, four of them on the Commerce Secretary’s own stage.

The buyer’s real reason

Countries other than the United States, she said, “may rationally decide to have another source of this technology,” not only for the military but for payment and credit-card systems. Her reason was that a state may negotiate that “the servers are in your country and the data centers are in your country and the data isn’t leaving” and still find that “if somebody somewhere else has a kill switch for your economy, that’s kind of bad.” A sovereign program’s envelope is not the same thing as its spending. A program that spends a fifth of one percent of its envelope in its first year has not left anyone’s landlord.

Chapter 13, A Change of Landlord. The sovereign buyers are the chapter’s third set of exits; the Gulf and orbit are the first two.