Sovereign AI: what it means in the US, UK, France, EU and the Gulf, and what is actually funded

Sovereign AI is a phrase every government now uses and no two governments mean the same thing by. In the United States it means exporting American chips and models to allies. In the UK it means a state-backed fund, a public supercomputer and a bargaining position. In France and the EU it means owning enough compute and cloud that a foreign switch cannot turn you off. In the Gulf it means buying the best hardware on earth with oil and sovereign-wealth money. This guide sorts what each place has actually paid for from what it has merely announced, as of October 6, 2026.

What sovereign AI actually means

The loosest definition comes from the seller. Nvidia's Jensen Huang, in a February 2024 post from the World Governments Summit, said every country needs to own the production of its own intelligence, which in practice means owning the data, the compute and the models. That is a sales pitch with a policy shape: the more countries accept it, the more of them buy accelerators.

The harder version is a dependency test. Can your hospitals, courts and defense ministries keep running AI services if a foreign provider, or a foreign export license, goes away? Under that test almost nobody is sovereign today, because almost everybody runs on Nvidia silicon. Keep that test in mind, because it is the one the headline figures below mostly fail to measure.

The United States: sovereignty as an export product

Washington's version is the mirror image of everyone else's. Executive Order 14320, signed July 23, 2025, tells Commerce, State and the science office to build an American AI Exports Program that packages chips, data centers, cloud, models and applications for foreign buyers. Its stated purpose is that American technologies, standards and governance models are adopted worldwide.

The order itself carries no dollar figure. The money is private. What the government does control is the license: on November 19, 2025, Commerce authorized Humain of Saudi Arabia and G42 of the UAE to buy up to 35,000 Blackwell-class GB300 chips each, conditioned on security and reporting requirements. So for the Gulf, the "sovereign" stack is sovereign inside a US permission slip.

The UK: one finished machine, one new fund

The UK has the cleanest record of money that exists. Isambard-AI at the University of Bristol launched on July 17, 2025, at a cost of £225 million, with 5,448 Nvidia GH200 Grace Hopper chips, and the university ranks it 11th fastest supercomputer in the world at launch. That is funded, built and running. The government's own one-year review of its AI Opportunities Action Plan, published January 29, 2026, also reports £28.2 billion of investment tied to five AI Growth Zones, a figure to treat as investor intent rather than spending, and earmarks up to £250 million for cloud capacity.

The Sovereign AI Unit is the new piece. The January review said the government would launch its next phase in April 2026, and on April 16, 2026 Technology Secretary Liz Kendall did, with up to £500 million to invest in UK AI companies. Note the word "up to": it is a ceiling for equity and support, not a cash pile that has been spent. Nvidia separately said in June 2026 that it is putting £2 billion into the UK startup ecosystem; that is a vendor commitment, not public money. For the wider ledger, see our earlier piece on what the UK has funded versus announced.

France: the biggest number, the thinnest paper

At the February 2025 Paris summit the Élysée said more than €109 billion in infrastructure investment had been announced. The government page gives that single aggregate. It gives no split by investor, no capacity in gigawatts and no delivery dates. Without a split we cannot say how much is public, private or foreign.

That matters for the sovereignty claim. A data center in France filled with American chips is French real estate and French electricity. It is not obviously French control. We could not find an official breakdown of how much of the €109 billion has been committed to construction, so treat it as an announcement.

The EU: €20 billion on paper, no winners yet

Brussels has the most structured process. The InvestAI initiative set aside €20 billion for AI gigafactories, each meant to run roughly 100,000 advanced chips, about four times a current AI factory, as the European Investment Bank and the Commission restated on December 4, 2025. The EuroHPC Joint Undertaking then published the formal call for tenders on July 30, 2026, with submissions due December 3, 2026.

So as of today the EU has funded a competition, not a gigafactory. No consortium has been selected and none has a site under construction. The pledge is real, but the earliest it can turn into hardware is after an award decision that has not been made.

The Gulf: money and chips, delivery unconfirmed

Humain, the Saudi sovereign-wealth subsidiary, and Nvidia announced on May 13, 2025 a plan for up to 500 megawatts over five years, with a first phase of 18,000 GB300 systems. G42's Talal Al Kaissi said in October 2025 that the first 200 megawatts of Stargate UAE would come online in 2026. Neither of those pages is a delivery notice. In the pages we opened, we found no confirmation that either first phase is energized, and a July 2026 Rest of World report noted that even Gulf states with billions to spend still depend on Nvidia for training chips.

Funded versus announced, side by side

• UK: funded and running, Isambard-AI at £225 million; announced as a ceiling, up to £500 million for the Sovereign AI Unit.
• US: no appropriation in the order; a licensing decision, up to 35,000 chips each for two Gulf firms.
• France: €109 billion announced in aggregate, with no public breakdown.
• EU: €20 billion earmarked, tender open until December 3, 2026, no awards.
• Gulf: up to 500 megawatts and a 200 megawatt first phase, targeted but not confirmed as delivered in our sources.

Our take

Sovereign AI is a label that works for whoever is selling: chipmakers sell it, ministers launch it, and the figures are rarely audited. The only country in this survey with a finished, public, verifiable machine is the UK, and at 5,448 chips it is small next to a single gigafactory specification. We would trust a project when it shows a site, a power connection and a date, and we would read every "up to" as a ceiling. And no government in this list has solved the dependency test: all of them still run on one American supplier.